BAYESIAN OPTIMIZATION

ROMAN GARNETT
BAYESIAN OPTIMIZATION

ROMAN GARNETT

Cambridge University Press

CON TEN TS

preface ix

notation xiii

1 introduction 1
1.1 Formalization of optimization 2
1.2 The Bayesian approach 5

2 gaussian processes 15
2.1 Definition and basic properties 16
2.2 Inference with exact and noisy observations 18
2.3 Overview of remainder of chapter 26
2.4 Joint Gaussian processes 26
2.5 Continuity 28
2.6 Differentiability 30
2.7 Existence and uniqueness of global maxima 33
2.8 Inference with non-Gaussian observations and constraints 35
2.9 Summary of major ideas 41

3 modeling with gaussian processes 45

3.1 The prior mean function 46
3.2 The prior covariance function 49
3.3 Notable covariance functions 51
3.4 Modifying and combining covariance functions 54
3.5 Modeling functions on high-dimensional domains 61
3.6 Summary of major ideas 64

4 model assessment, selection, and averaging 67

4.1 Models and model structures 68
4.2 Bayesian inference over parametric model spaces 70
4.3 Model selection via posterior maximization 73
4.4 Model averaging 74
4.5 Multiple model structures 78
4.6 Automating model structure search 81
4.7 Summary of major ideas 84

5 decision theory for optimization 87

5.1 Introduction to Bayesian decision theory 89
5.2 Sequential decisions with a fixed budget 91
5.3 Cost and approximation of the optimal policy 99
5.4 Cost-aware optimization and termination as a decision 103
5.5 Summary of major ideas 106

6 utility functions for optimization 109

6.1 Expected utility of terminal recommendation 109
6.2 Cumulative reward 114

This material will be published by Cambridge University Press as Bayesian Optimization. v

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
contents

6.3 Information gain 115

6.4 Dependence on model of objective function 116
6.5 Comparison of utility functions 117
6.6 Summary of major ideas 119

7 common bayesian optimization policies 123

7.1 Example optimization scenario 124
7.2 Decision-theoretic policies 124
7.3 Expected improvement 127
7.4 Knowledge gradient 129
7.5 Probability of improvement 131
7.6 Mutual information and entropy search 135
7.7 Multi-armed bandits and optimization 141
7.8 Maximizing a statistical upper bound 145
7.9 Thompson sampling 148
7.10 Other ideas in policy construction 150
7.11 Summary of major ideas 156

8 computing policies with gaussian processes 157

8.1 Notation for objective function model 157
8.2 Expected improvement 158
8.3 Probability of improvement 167
8.4 Upper confidence bound 170
8.5 Approximate computation for one-step lookahead 171
8.6 Knowledge gradient 172
8.7 Thompson sampling 176
8.8 Mutual information with 𝑥 ∗ 180
8.9 Mutual information with 𝑓 ∗ 187
8.10 Averaging over a space of Gaussian processes 192
8.11 Alternative models: Bayesian neural networks, etc. 196
8.12 Summary of major ideas 200

9 implementation 201
9.1 Gaussian process inference, scaling, and approximation 201
9.2 Optimizing acquisition functions 207
9.3 Starting and stopping optimization 210
9.4 Summary of major ideas 212

10 theoretical analysis 213

10.1 Regret 213
10.2 Useful function spaces for studying convergence 215
10.3 Relevant properties of covariance functions 220
10.4 Bayesian regret with observation noise 224
10.5 Worst-case regret with observation noise 232
10.6 The exact observation case 237
10.7 The effect of unknown hyperparameters 241
10.8 Summary of major ideas 243

vi
 contents

11 extensions and related settings 245

11.1 Unknown observation costs 245
11.2 Constrained optimization and unknown constraints 249
11.3 Synchronous batch observations 252
11.4 Asynchronous observation with pending experiments 262
11.5 Multifidelity optimization 263
11.6 Multitask optimization 266
11.7 Multiobjective optimization 269
11.8 Gradient observations 276
11.9 Stochastic and robust optimization 277
11.10 Incremental optimization of sequential procedures 281
11.11 Non-Gaussian observation models and active search 282

12 a brief history of bayesian optimization 287

12.1 Historical precursors and optimal design 287
12.2 Sequential analysis and Bayesian experimental design 287
12.3 The rise of Bayesian optimization 289
12.4 Later rediscovery and development 290
12.5 Multi-armed bandits to infinite-armed bandits 292
12.6 What’s next? 294

a the gaussian distribution 295

b methods for approximate bayesian inference 301

c gradients 307

d annotated bibliography of applications 313

bibliography 331

index 353

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ vii

PREFACE

My interest in Bayesian optimization began in 2007 at the start of my

doctoral studies. I was frustrated that there seemed to be a Bayesian
approach to every task I cared about, except optimization. Of course, as
was often the case at that time (not to mention now!), I was mistaken in
this belief, but one should never let ignorance impede inspiration.
Meanwhile, my labmate and soon-to-be frequent collaborator Mike
Osborne had a fresh copy of rasmussen and williams’s Gaussian Pro-
cesses for Machine Learning and just would not stop talking about gps at
our lab meetings. Through sheer brute force of repetition, I slowly built
a hand-wavy intuition for Gaussian processes – my mental model was
the “sausage plot” – without even being sure about their precise defi-
nition. However, I was pretty sure that marginals were Gaussian (what
else?), and one day it occurred to me that one could achieve Bayesian The first of many “sausage plots” to come.
optimization by maximizing the probability of improvement. This was
the algorithm I was looking for! In my excitement I shot off an email to
Mike that kicked off years of fruitful collaboration:

Can I ask a dumb question about gps? Let’s say that I’m doing
function approximation on an interval with a gp. So I’ve got this
mean function 𝑚(𝑥) and a variance function

𝑣 (𝑥). Is it true

that if
I pick a particular point 𝑥, then 𝑝 𝑓 (𝑥) ∼ N 𝑚(𝑥), 𝑣 (𝑥) ? Please
say yes.
If this is true, then I think the idea of doing Bayesian optimization
using gps is, dare I say, trivial.

The hubris of youth!

Well, it turned out I was 45 years too late in proposing this algo-
rithm,1 and that it only seemed “trivial” because I had no appreciation for 1 h. j. kushner (1962). A Versatile Stochastic
its theoretical foundation. However, truly great ideas are rediscovered Model of a Function of Unknown and Time
Varying Form. Journal of Mathematical Analy-
many times, and my excitement did not fade. Once I developed a deeper sis and Applications 5(1):150–167.
understanding of Gaussian processes and Bayesian decision theory, I
came to see them as a “Bayesian crank” I could turn to realize adaptive
algorithms for any task. I have been repeatedly astonished to find that
the resulting algorithms – seemingly by magic – automatically display
intuitive emergent behavior as a result of their careful design. My goal
with this book is to paint this grand picture. In effect, it is a gift to my
former self: the book I wish I had in the early years of my career.
In the context of machine learning, Bayesian optimization is an
ancient idea – kushner’s paper appeared only three years after the
term “machine learning” was coined! Despite its advanced age, Bayesian
optimization has been enjoying a period of revitalization and rapid
progress over the past ten years. The primary driver of this renaissance
has been advances in computation, which have enabled increasingly
sophisticated tools for Bayesian modeling and inference.
Ironically, however, perhaps the most critical development was not
Bayesian at all, but the rise of deep neural networks, another old idea

This material will be published by Cambridge University Press as Bayesian Optimization. ix

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 preface

granted new life by modern computation. The extreme cost of training

these models demands efficient routines for hyperparameter tuning, and
in a timely and influential paper, snoek et al. demonstrated (dramati-
2 j. snoek et al. (2012). Practical Bayesian Op- cally!) that Bayesian optimization was up to the task.2 Hyperparameter
timization of Machine Learning Algorithms. tuning proved to be a “killer app” for Bayesian optimization, and the en-
suing surge of interest has yielded a mountain of publications developing
neurips 2012.

new algorithms and improving old ones, exploring countless variations

on the basic setup, establishing theoretical guarantees on performance,
and applying the framework to a huge range of domains.
Due to the nature of the computer science publication model, these
recent developments are scattered across dozens of brief papers, and the
pressure to establish novelty in a limited space can obscure the big picture
in favor of minute details. This book aims to provide a self-contained and
comprehensive introduction to Bayesian optimization, starting “from
scratch” and carefully developing all the key ideas along the way. This
bottom-up approach allows us to identify unifying themes in Bayesian
optimization algorithms that may be lost when surveying the literature.
intended audience The intended audience is graduate students and researchers in ma-
chine learning, statistics, and related fields. However, it is also my sincere
hope that practitioners from more distant fields wishing to harness the
power of Bayesian optimization will also find some utility here.
prerequisites For the bulk of the text, I assume the reader is comfortable with differ-
ential and integral calculus, probability, and linear algebra. On occasion
the discussion will meander to more esoteric areas of mathematics, and
these passages can be safely ignored and returned to later if desired. A
good working knowledge of the Gaussian distribution is also essential,
and I provide an abbreviated but sufficient introduction in appendix a.
chapters 2–4: modeling the objective function The book is divided into three main parts. Chapters 2–4 cover theo-
with Gaussian processes retical and practical aspects of modeling with Gaussian processes. This
class of models is the overwhelming favorite in the Bayesian optimiza-
tion literature, and the material contained within is critical for several
following chapters. It was daunting to write this material in light of
the many excellent references already available, in particular the afore-
mentioned Gaussian Processes for Machine Learning. However, I heavily
biased the presentation in light of the needs of optimization, and even
experts may find something new.
chapters 5–7: sequential decision making and Chapters 5–7 develop the theory of sequential decision making and
policy building its application to optimization. Although this theory requires a model
of the objective function and our observations of it, the presentation is
agnostic to the choice of model and may be read independently from the
preceding chapters on Gaussian processes.
chapters 8–10: Bayesian optimization with These threads are unified in chapters 8–10, which discuss the partic-
Gaussian processes ulars of Bayesian optimization with Gaussian process models. Chapters
8–9 cover details of computation and implementation, and chapter 10
discusses theoretical performance bounds on Bayesian optimization al-
gorithms, where most results depend intimately on a Gaussian process
model of the objective function or the associated reproducing kernel
Hilbert space.

x
 preface

The nuances of some applications require modifications to the basic chapter 11: extensions
sequential optimization scheme that is the focus of the bulk of the book,
and chapter 11 introduces several notable extensions to this basic setup.
Each is systematically presented through the unifying lens of Bayesian
decision theory to illustrate how one might proceed when facing a novel
situation.
Finally, chapter 12 provides a brief and standalone history of Bayesian chapter 12: brief history of Bayesian
optimization. This was perhaps the most fun chapter for me to write, optimization
if only because it forced me to plod through old Soviet literature (in an
actual library! what a novelty these days!). To my surprise I was able to
antedate many Bayesian optimization policies beyond their commonly
attested origin, including expected improvement, knowledge gradient,
probability of improvement, and upper confidence bound. (A reader
familiar with the literature may be surprised to learn the last of these
was actually the first policy discussed by kushner in his 1962 paper.)
Despite my best efforts, there may still be stones left to be overturned
before the complete history is revealed.
Dependencies between the main chapters are illustrated in the mar-
gin. There are two natural linearizations of the material. The first is the 2 3 4
one I adopted and personally prefer, which covers modeling prior to
decision making. However, one could also proceed in the other order,
7 6 5
reading chapters 5–7 first, then looping back to chapter 2. After covering
the material in these chapters (in either order), the remainder of the book
can be perused at will. Logical partial paths through the book include:
8
• a minimal but self-contained introduction: chapters 1–2, 5–7
• a shorter introduction requiring leaps of faith: chapters 1 and 7
9
• a crash course on the underlying theory: chapters 1–2, 5–7, 10
• a head start on implementing a software package: chapters 1–9
10
A reader already quite comfortable with Gaussian processes might wish
to skip over chapters 2–4 entirely.
I struggled for some time over whether to include a chapter on ap- 11
plications. On the one hand, Bayesian optimization ultimately owes its
popularity to its success in optimizing a growing and diverse set of dif-
A dependency graph for chapters 2–11. Chap-
ficult objectives. However, these applications often require extensive ter 1 is a universal dependency.
technical background to appreciate, and an adequate coverage would
be tedious to write and tedious to read. As a compromise, I provide an
annotated bibliography outlining the optimization challenges involved annotated bibliography of applications:
in notable domains of interest and pointing to studies where these chal- appendix d, p. 313
lenges were successfully overcome with the aid of Bayesian optimization.
The sheer size of the Bayesian optimization literature – especially
the output of the previous decade – makes it impossible to provide a
complete survey of every recent development. This is especially true
for the extensions discussed in chapter 11 and even more so for the
bibliography on applications, where work has proliferated in myriad
branching directions. Instead I settled for presenting what I considered
to be the most important ideas and providing pointers to entry points

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ xi

preface

for the relevant literature. The reader should not read anything into any
omissions; there is simply too much high-quality work to go around.
Additional information about the book, including a list of errata as
they are discovered, may be found at the companion webpage:

bayesoptbook.com

I encourage the reader to report any errata or other issues to the com-
panion GitHub repository for discussion and resolution:

github.com/bayesoptbook/bayesoptbook.github.io

Thank you! Preparation of this manuscript was facilitated tremendously by nu-

merous free and open source projects, and the creators, developers, and
maintainers of these projects have my sincere gratitude. The manuscript
was typeset in LATEX using the excellent and extremely flexible memoir
class. The typeface is Linux Libertine. Figures were laid out in matlab
and converted to TikZ/pgf/pgfplots for further tweaking and typeset-
ting via the matlab2tikz script. The colors used in figures were based
on www.colorbrewer.org by Cynthia A. Brewer, and I endeavored to the
best of my ability to ensure that the figures are colorblind friendly. The
colormap used in heat maps is a slight modification of the Matplotlib
viridis colormap where the “bright” end is pure white.
I would like to thank Eric Brochu, Nando de Freitas, Matt Hoffman,
Frank Hutter, Mike Osborne, Bobak Shahriari, Jasper Snoek, Kevin Swer-
sky, and Ziyu Wang, who jointly provided the activation energy for this
undertaking. I would also like to thank Eytan Bakshy, Ivan Barrientos,
George De Ath, Peter Frazier, Lukas Fröhlich, Ashok Gautam, Jake Gard-
ner, Javier González, Eugen Hotaj, Frank Hutter, Jungtaek Kim, Bryan
Low, Ruben Martinez-Cantin, Keita Mori, Kevin Murphy, Mike Osborne,
Matthias Poloczeck, Jon Scarlett, Bobak Shahriari, Jasper Snoek, Sebas-
tian Tay, Sattar Vakili, Qiuyi Zhang, and GitHub users cgoble001 and
chaos-and-patterns for their suggestions, corrections, and valuable
discussions along the way, as well as David Tranah and Anna Scriven
at Cambridge University Press for their support and patience as I con-
tinually missed deadlines. Special thanks are due to the students of two
seminars run at Washington University covering the material in this
book; their feedback was also instrumental in shaping the book’s content.
Funding support was provided by the United States National Science
Foundation (nsf) under award number 1845434. Any opinions, findings,
and conclusions or recommendations expressed in this book are those
of the author and do not necessarily reflect the views of the nsf.
This book took far more time than I initially anticipated, and I would
especially like to thank my wife Marion and son Max (arg Max?) for their
understanding and support during this long journey.

Roman Garnett
St. Louis, Missouri
January 2022

xii
NOTATION

All vectors are column vectors and are denoted in lowercase bold: x ∈ ℝ𝑑. vectors and matrices
Matrices are denoted in uppercase bold: A.
We adopt the “numerator layout” convention for matrix calculus: matrix calculus convention
the derivative of a vector by a scalar is a (column) vector, whereas the
derivative of a scalar by a vector is a row vector. This results in the chain
rule proceeding from left-to-right; for example, if a vector x(𝜃 ) depends chain rule
on a scalar parameter 𝜃 , then for a function 𝑓 (x), we have:

𝜕𝑓 𝜕𝑓 𝜕x
= .
𝜕𝜃 𝜕x 𝜕𝜃

When an indicator function is required, we use the Iverson bracket indicator functions
notation. For a statement 𝑠, we have:
(
1 if 𝑠 is true;
[𝑠] =
0 otherwise.

The statement may depend on a parameter: [𝑥 ∈ 𝐴], [𝑥 ≥ 0], etc.

Logarithms are taken with respect to their natural base, 𝑒. Quantities logarithms
in log units such as log likelihoods or entropy thus have units of nats, nats
the base-𝑒 analogue of the more familiar base-2 bits.

symbols with implicit dependence on location

There is one notational innovation in this book compared with the
Gaussian process and Bayesian optimization literature at large: we make
heavy use of symbols for quantities that depend implicitly on a putative
(and arbitrary) input location 𝑥. Most importantly, to refer to the value
of an objective function 𝑓 at a given location 𝑥, we introduce the symbol
𝜙 = 𝑓 (𝑥). This avoids clash with the name of the function itself, 𝑓, while
avoiding an extra layer of brackets. We use this scheme throughout the
book, including variations such as:

𝜙 0 = 𝑓 (𝑥 0); 𝝓 = 𝑓 (x); 𝛾 = 𝑔(𝑥); etc.

To refer to the outcome of a (possibly inexact) measurement at 𝑥, we use

the symbol 𝑦; the distribution of 𝑦 presumably depends on 𝜙.
We also allocate symbols to describe properties of the marginal pre-
dictive distributions for the objective function value 𝜙 and observed
value 𝑦, all of which also have implicit dependence on 𝑥. These appear
in the table below.

comprehensive list of symbols

A list of important symbols appears on the following pages, arranged
roughly in alphabetical order.

This material will be published by Cambridge University Press as Bayesian Optimization. xiii
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
notation

symbol description
≡ identical equality of functions; for a constant 𝑐, 𝑓 ≡ 𝑐 is a constant function
∇ gradient operator
∅ termination option: the action of immediately terminating optimization
≺ either Pareto dominance or the Löwner order: for symmetric A, B, A ≺ B if and only if B − A
is positive definite
is sampled according to: 𝜔 is a realization of a random
variable with probability density 𝑝 (𝜔)
à à Р
𝜔 ∼ 𝑝 (𝜔)
𝑖 X𝑖 disjoint union of {X𝑖 }: 𝑖 X𝑖 = 𝑖 (𝑥, 𝑖) | 𝑥 ∈ X𝑖
|A| determinant of square matrix A
|x| Euclidean norm of vector x; |x − y| is thus the Euclidean distance between vectors x and y
k𝑓 k H𝐾 norm of function 𝑓 in reproducing kernel Hilbert space H𝐾
A−1 inverse of square matrix A
x> transpose of vector x
0 vector or matrix of zeros
A action space for a decision
𝛼 (𝑥; D) acquisition function evaluating 𝑥 given data D
𝛼𝜏 (𝑥; D) expected marginal gain in 𝑢 (D) after observing at 𝑥 then making 𝜏 − 1 additional optimal
observations given the outcome
𝛼𝜏∗ (D) value of D with horizon 𝜏: expected marginal gain in 𝑢 (D) from 𝜏 additional optimal obser-
vations
𝛼 ei expected improvement
𝛼𝑓 ∗ mutual information between 𝑦 and 𝑓 ∗
𝛼 kg knowledge gradient
𝛼 pi probability of improvement
𝛼𝑥 ∗ mutual information between 𝑦 and 𝑥 ∗
𝛼 ucb upper confidence bound
𝛼 ts Thompson sampling “acquisition function:” a draw 𝑓 ∼ 𝑝 (𝑓 | D)
𝛽 confidence parameter in Gaussian process upper confidence bound policy
𝛽 (x; D) batch acquisition function evaluating x given data D; may have modifiers analogous to 𝛼
C prior covariance matrix of observed values y: C = cov[y]
𝑐 (D) cost of acquiring data D
chol A Cholesky decomposition of positive definite matrix A: if 𝚲 = chol A, then A = 𝚲𝚲>
corr[𝜔,𝜓 ] correlation of random variables 𝜔 and 𝜓 ; with a single argument, corr[𝜔] = corr[𝜔, 𝜔]
cov[𝜔,𝜓 ] covariance of random variables 𝜔 and 𝜓 ; with a single argument, cov[𝜔] = cov[𝜔, 𝜔]
D set of observed data, D = (x, y) 
D,0 D1 set of observed data after observing at 𝑥: D 0 = D ∪ (𝑥, 𝑦) = (x,0 y0)
D𝜏 set of observed data after 𝜏 observations
𝐷 kl [𝑝 k 𝑞] Kullback–Leibler divergence between distributions with probability densities 𝑝 and 𝑞
Δ(𝑥, 𝑦) marginal gain in utility after acquiring observation (𝑥, 𝑦): Δ(𝑥, 𝑦) = 𝑢 (D 0) − 𝑢 (D)
𝛿 (𝜔 − 𝑎) Dirac delta distribution on 𝜔 with point mass at 𝑎
diag x diagonal matrix with diagonal x
𝔼, 𝔼𝜔 expectation, expectation with respect to 𝜔
𝜀 measurement error associated with an observation at 𝑥: 𝜀 = 𝑦 − 𝜙
𝑓 objective function; 𝑓 : X → ℝ
𝑓 |Y the restriction of 𝑓 onto the subdomain Y ⊂ X
𝑓∗ globally maximal value of the objective function: 𝑓 ∗ = max 𝑓
𝛾𝜏 information capacity of an observation process given 𝜏 iterations

xiv
 notation

symbol description
GP (𝑓 ; 𝜇, 𝐾) Gaussian process on 𝑓 with mean function 𝜇 and covariance function 𝐾
H𝐾 reproducing kernel Hilbert space associated with kernel 𝐾
H𝐾 [𝐵] ball of radius 𝐵 in H𝐾 : {𝑓 | k𝑓 k H𝐾 ≤ 𝐵}
𝐻 [𝜔] discrete or differential entropy of random variable 𝜔
𝐻 [𝜔 | D] discrete or differential of random variable 𝜔 after conditioning on D
𝐼 (𝜔;𝜓 ) mutual information between random variables 𝜔 and 𝜓
𝐼 (𝜔;𝜓 | D) mutual information between random variables 𝜔 and 𝜓 after conditioning on D
I identity matrix
𝐾 prior covariance function: 𝐾 = cov[𝑓 ]
𝐾D posterior covariance function given data D: 𝐾D = cov[𝑓 | D]
𝐾m Matérn covariance function
𝐾se squared exponential covariance function
𝜅 cross covariance between 𝑓 and observed values y: 𝜅 (𝑥) = cov[y, 𝜙 | 𝑥]
ℓ either a length-scale parameter or the lookahead horizon
𝜆 output-scale parameter
M space of models indexed by the hyperparameter vector 𝜽
m prior expected value of observed values y, m = 𝔼[y]
𝜇 either the prior mean function, 𝜇 = 𝔼[𝑓 ], or the predictive mean of 𝜙: 𝜇 = 𝔼[𝜙 | 𝑥, D] = 𝜇D (𝑥)
𝜇D posterior mean function given data D: 𝜇D = 𝔼[𝑓 | D]
N (𝝓; 𝝁, 𝚺) multivariate normal distribution on 𝝓 with mean vector 𝝁 and covariance matrix 𝚺
N measurement error covariance corresponding to observed values y
O is asymptotically bounded above by: for nonnegative functions 𝑓, 𝑔 of 𝜏, 𝑓 = O(𝑔) if 𝑓 /𝑔 is
asymptotically bounded by a constant as 𝜏 → ∞
O∗ as above with logarithmic factors suppressed: 𝑓 = O∗ (𝑔) if 𝑓 (𝜏) (log 𝜏)𝑘 = O(𝑔) for some 𝑘
Ω is asymptotically bounded below by: 𝑓 = Ω(𝑔) if 𝑔 = O(𝑓 )
𝑝 probability density
𝑞 either an approximation to probability density 𝑝 or a quantile
∫𝑧 function
Φ(𝑧) standard normal cumulative density function: Φ(𝑧) = −∞ 𝜙 (𝑧 0) d𝑧 0
𝜙 value of the objective function at 𝑥: 𝜙 = 𝑓 (𝑥) √
𝜙 (𝑧) standard normal probability density function: 𝜙 (𝑧) = ( 2𝜋) −1 exp(− 21 𝑧 2 )
Pr probability
ℝ set of real numbers
𝑅𝜏 cumulative regret after 𝜏 iterations
𝑅¯𝜏 [𝐵] worst-case cumulative regret after 𝜏 iterations on the rkhs ball H𝐾 [𝐵]
𝑟𝜏 simple regret after 𝜏 iterations
𝑟¯𝜏 [𝐵] worst-case simple regret after 𝜏 iterations on the rkhs ball H𝐾 [𝐵]
P a correlation matrix
𝜌 a scalar correlation
𝜌𝜏 instantaneous regret on iteration 𝜏
𝑠2 predictive variance of 𝑦; for additive Gaussian noise, 𝑠 2 = var[𝑦 | 𝑥, D] = 𝜎 2 + 𝜎𝑛2
𝚺 a covariance matrix, usually the Gram matrix associated with x: 𝚺 = 𝐾D (x, x)
𝜎2 predictive variance of 𝜙: 𝜎 2 = 𝐾D (𝑥, 𝑥)
𝜎𝑛2 variance of measurement error at 𝑥: 𝜎𝑛2 = var[𝜀 | 𝑥]
std[𝜔] standard deviation of random variable 𝜔
T (𝜙; 𝜇, 𝜎 ,2 𝜈) Student-𝑡 distribution on 𝜙 with 𝜈 degrees of freedom, mean 𝜇, and variance 𝜎 2
TN (𝜙; 𝜇, 𝜎 ,2 𝐼 ) truncated normal distribution, N (𝜙; 𝜇, 𝜎 2 ) truncated to interval 𝐼

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ xv

notation

symbol description
𝜏 either decision horizon (in the context of decision making) or number of optimization itera-
tions passed (in the context of asymptotic analysis)
Θ is asymptotically bounded above and below by: 𝑓 = Θ(𝑔) if 𝑓 = O(𝑔) and 𝑓 = Ω(𝑔)
𝜽 vector of hyperparameters indexing a model space M
tr A trace of square matrix A
𝑢 (D) utility of data D
var[𝜔] variance of random variable 𝜔
𝑥 putative input location of the objective function
x either a sequence of observed locations x = {𝑥𝑖 } or (when the distinction is important) a
vector-valued input location
𝑥∗ a location attaining the globally maximal value of 𝑓 : 𝑥 ∗ ∈ arg max 𝑓 ; 𝑓 (𝑥 ∗ ) = 𝑓 ∗
X domain of objective function
𝑦 value resulting from an observation at 𝑥
y observed values resulting from observations at locations x
𝑧 𝑧-score of measurement 𝑦 at 𝑥: 𝑧 = (𝑦 − 𝜇)/𝑠

xvi
 1
IN TRODUCTION

Optimization is an innate human behavior. On an individual level, we

all strive to better ourselves and our surroundings. On a collective level,
societies struggle to allocate limited resources seeking to improve the
welfare of their members, and optimization has been an engine of societal
progress since the domestication of crops through selective breeding
over 12 000 years ago – an effort that continues to this day.
Given its pervasiveness, it should perhaps not be surprising that
optimization is also difficult. While searching for an optimal design,
we must spend – sometimes quite significant – resources evaluating
suboptimal alternatives along the way. This observation compels us to
seek methods of optimization that, when necessary, can carefully allocate
resources to identify optimal parameters as efficiently as possible. This
is the goal of mathematical optimization.
Since the 1960s, the statistics and machine learning communities
have refined a Bayesian approach to optimization that we will develop
and explore in this book. Bayesian optimization routines rely on a statis-
tical model of the objective function, whose beliefs guide the algorithm
in making the most fruitful decisions. These models can be quite so-
phisticated, and maintaining them throughout optimization may entail
significant cost of its own. However, the reward for this effort is unparal-
leled sample efficiency. For this reason, Bayesian optimization has found
a niche in optimizing objectives that:
• are costly to compute, precluding exhaustive evaluation,
• lack a useful expression, causing them to function as “black boxes,”
• cannot be evaluated exactly, but only through some indirect or noisy
mechanism, and/or
• offer no efficient mechanism for estimating their gradient.
Let us consider an example setting motivating the machine learn-
ing community’s recent interest in Bayesian optimization. Consider a
data scientist crafting a complex machine learning model – say a deep
neural network – from training data. To ensure success, the scientist
must carefully tune the model’s hyperparameters, including the network
architecture and details of the training procedure, which have massive
influence on performance. Unfortunately, effective settings can only be
identified via trial-and-error: by training several networks with different
settings and evaluating their performance on a validation dataset.
The search for the best hyperparameters is of course an exercise
in optimization. Mathematical optimization has been under continual
development for centuries, and numerous off-the-shelf procedures are
available. However, these procedures usually make assumptions about
the objective function that may not always be valid. For example, we
might assume that the objective is cheap to evaluate, that we can easily
compute its gradient, or that it is convex, allowing us to reduce from
global to local optimization.

This material will be published by Cambridge University Press as Bayesian Optimization. 1

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 introduction

In hyperparameter tuning, all of these assumptions are invalid. Train-

ing a deep neural network can be extremely expensive in terms of both
time and energy. When some hyperparameters are discrete – as many
features of network architecture naturally are – the gradient does not
even exist. Finally, the mapping from hyperparameters to performance
may be highly complex and multimodal, so local refinement may not
yield an acceptable result.
The Bayesian approach to optimization allows us to relax all of these
1 r. turner et al. (2021). Bayesian Optimization assumptions when necessary, and Bayesian optimization algorithms can
is Superior to Random Search for Machine deliver impressive performance even when optimizing complex “black
Learning Hyperparameter Tuning: Analysis of
the Black-Box Optimization Challenge 2020.
box” objectives under severely limited observation budgets. Bayesian
Proceedings of the Neurips 2020 Competition optimization has proven successful in settings spanning science, engi-
and Demonstration Track. neering, and beyond, including of course hyperparameter tuning.1 In
2 a. gelman and a. vehtari (2021). What are light of this broad success, gelman and vehtari identified adaptive
the most important statistical ideas of the past decision analysis – and Bayesian optimization in particular – as one of
50 years? Journal of the American Statistical
Association.
the eight most important statistical ideas of the past 50 years.2
Covering all these applications and their nuances could easily fill a
separate volume (although we do provide an overview of some impor-
annotated bibliography of applications: tant application domains in an annotated bibliography), so in this book
appendix d, p. 313 we will settle for developing the mathematical foundation of Bayesian
optimization underlying its success. In the remainder of this chapter we
will lay important groundwork for this discussion. We will first establish
the precise formulation of optimization we will consider and important
conventions of our presentation, then outline and illustrate the key as-
outline and reading guide: p. x pects of the Bayesian approach. The reader may find an outline of and
reading guide for the chapters to come in the preface.

1.1 formalization of optimization

Throughout this book we will consider a simple but flexible formulation
of sequential global optimization outlined below. There is nothing inher-
ently Bayesian about this model, and countless solutions are possible.
objective function, 𝑓 We begin with a real-valued objective function defined on some
domain of objective function, X domain X ; 𝑓 : X → ℝ. We make no assumptions regarding the nature
of the domain. In particular, it need not be Euclidean but might instead,
𝑓∗ for example, comprise a space of complex structured objects. The goal
of optimization is to systematically search the domain for a point 𝑥 ∗
attaining the globally maximal value 𝑓 ∗ :3

𝑥 ∗ ∈ arg max 𝑓 (𝑥); 𝑓 ∗ = max 𝑓 (𝑥) = 𝑓 (𝑥 ∗ ). (1.1)

𝑥 ∈X 𝑥 ∈X
𝑥∗
Before we proceed, we note that our focus on maximization rather
An objective function with the location, 𝑥 ,∗
and value, 𝑓 ,∗ of the global optimum marked.
than minimization is entirely arbitrary; the author simply judges max-
imization to be the more optimistic choice. If desired, we can freely
transform one problem to the other by negating the objective function.
3 A skeptical reader may object that, without
further assumptions, a global maximum may We caution the reader that some translation may be required when com-
not exist at all! We will sidestep this issue for paring expressions derived here to what may appear in parallel texts
now and pick it up again in § 2.7, p. 34. focusing on minimization.

2
 1.1. formalization of optimization

input: initial dataset D I can be empty Algorithm 1.1: Sequential optimization.

repeat
𝑥 ← policy(D) I select the next observation location
𝑦 ← observe(𝑥)
 I observe at the chosen location
D ← D ∪ (𝑥, 𝑦) I update dataset
until termination condition reached I e.g., budget exhausted
return D

In a significant departure from classical mathematical optimization,

we do not require that the objective function have a known functional
form or even be computable directly. Rather, we only require access to a
mechanism revealing some information about the objective function at
identified points on demand. By amassing sufficient information from
this mechanism, we may hope to infer the solution to (1.1). Avoiding
the need for an explicit expression for 𝑓 allows us to consider so-called 4 Of course, we do not require but merely allow
“black box” optimization, where a system is optimized through indirect that the objective function act as a black box.
Access to a closed-form expression does not
measurements of its quality. This is one of the greatest strengths of preclude a Bayesian approach!
Bayesian optimization.4

Optimization policy
Directly solving for the location of global optima is infeasible except in
exceptional circumstances. The tools of traditional calculus are virtually
powerless in this setting; for example, enumerating and classifying every
stationary point in the domain would be tedious at best and perhaps
even impossible. Mathematical optimization instead takes an indirect
approach: we design a sequence of experiments to probe the objective
function for information that, we hope, will reveal the solution to (1.1).
The iterative procedure in algorithm 1.1 formalizes this process. We
begin with an initial (possibly empty) dataset D that we grow incremen-
tally through a sequence of observations of our design. In each iteration,
an optimization policy inspects the available data and selects a point
𝑥 ∈ X where we make our next observation.5 This action in turn reveals 5 Here “policy” has the same meaning as in
a corresponding value 𝑦 provided by the system under study. We append other decision-making contexts: it maps our
state (indexed by our data, D) to an action (the
the newly observed information to our dataset and finally decide whether location of our next observation, 𝑥).
to continue with another observation or terminate and return the current
data. When we inevitably do choose to terminate, the returned data can
be used by an external consumer as desired, for example to inform a
subsequent decision. terminal recommendations: § 5.1, p. 90
We place no restrictions on how an optimization policy is imple-
mented beyond mapping an arbitrary dataset to some point in the do-
main for evaluation. A policy may be deterministic or stochastic, as
demonstrated respectively by the prototypical examples of grid search
and random search. In fact, these popular policies are nonadaptive and
completely ignore the observed data. However, when observations only
come at significant cost, we will naturally prefer policies that adapt their
behavior in light of evolving information. The primary challenge in opti-

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 3

introduction

mization is designing policies that can rapidly optimize a broad class of

objective functions, and intelligent policy design will be our focus for
the majority of this book.

Observation model
For optimization to be feasible, the observations we obtain must provide
information about the objective function that can guide our search and in
aggregate determine the solution to (1.1). A near-universal assumption in
mathematical optimization is that observations yield exact evaluations of
the objective function at our chosen locations. However, this assumption
is unduly restrictive: many settings feature inexact measurements due
to noisy sensors, imperfect simulation, or statistical approximation. A
typical example featuring additive observation noise is shown in the
Inexact observations of an objective function margin. Although the objective function is not observed directly, the
corrupted by additive noise. noisy measurements nonetheless constrain the plausible options due to
strong dependence on the objective.
We thus relax the assumption of exact observation and instead as-
sume that observations are realized by a stochastic mechanism depending
measured value, 𝑦 on the objective function. Namely, we assume that the value 𝑦 resulting
observation location, 𝑥 from an observation at some point 𝑥 is distributed according to an ob-
servation model depending on the underlying objective function value
objective function value, 𝜙 = 𝑓 (𝑥) 𝜙 = 𝑓 (𝑥):
𝑝 (𝑦 | 𝑥, 𝜙). (1.2)
Through judicious design of the observation model, we may consider a
wide range of observation mechanisms.
conditional independence of observations As with the optimization policy, we do not make any assumptions
given objective values about the nature of the observation model, save one. Unless otherwise
mentioned, we assume that a set of multiple measurements y are condi-
tionally independent given the corresponding observation locations x
and objective function values 𝝓 = 𝑓 (x):
Ö
𝑝 (y | x, 𝝓) = 𝑝 (𝑦𝑖 | 𝑥𝑖 , 𝜙𝑖 ). (1.3)
𝑖
𝑝 (𝑦 | 𝑥, 𝜙)
This is not strictly necessary but is overwhelmingly common in practice
and will simplify our presentation considerably.
One particular observation model will enjoy most of our attention in
𝑦 this book: additive Gaussian noise. Here we model the value 𝑦 observed
at 𝑥 as
𝜙 𝑦 = 𝜙 + 𝜀,
Additive Gaussian noise: the distribution where 𝜀 represents measurement error. Errors are assumed to be Gaussian
of the value 𝑦 observed at 𝑥 is Gaussian, distributed with mean zero, implying a Gaussian observation model:
centered on the objective function value 𝜙.
𝑝 (𝑦 | 𝑥, 𝜙, 𝜎𝑛 ) = N (𝑦; 𝜙, 𝜎𝑛2 ). (1.4)

observation noise scale, 𝜎𝑛 Here the observation noise scale 𝜎𝑛 may optionally depend on 𝑥, allowing
heteroskedastic noise: § 2.2, p. 25 us to model both homoskedastic or heteroskedastic errors.

4
 1.2. the bayesian approach

If we take the noise scale to be identically zero, we recover the 𝑝 (𝑦 | 𝜙)

special case of exact observation, where we simply have 𝑦 = 𝜙 and the
observation model collapses to a Dirac delta distribution:
𝑝 (𝑦 | 𝜙) = 𝛿 (𝑦 − 𝜙).
Although not universally applicable, many settings do feature exact 𝑦
observations such as optimizing the output of a deterministic computer 𝜙
simulation. We will sometimes consider the exact case separately as
some results simplify considerably in the absence of measurement error. Exact observations: every value measured
We will focus on additive Gaussian noise as it is a reasonably faithful equals the corresponding function value,
yielding a Dirac delta observation model.
model for many systems and offers considerable mathematical conve-
nience. This observation model will be most prevalent in our discussion
on Gaussian processes in the next three chapters and on the explicit
computation of Bayesian optimization policies with this model class inference with non-Gaussian observations:
in chapter 8. However, the general methodology we will build in the § 2.8, p. 35
remainder of this book is not contingent on this choice, and we will optimization with non-Gaussian
occasionally address alternative observation mechanisms. observations: § 11.11, p. 282

Termination
The final decision we make in each iteration of optimization is whether
to terminate immediately or continue with another observation. As with
the optimization policy, we do not assume any particular mechanism by
which this decision is made. Termination may be deterministic – such
as stopping after reaching a certain optimization goal or exhausting
a preallocated observation budget – or stochastic, and may optionally
depend on the observed data. In many cases, the time of termination
may in fact not be under the control of the optimization routine at all
but instead decided by an external agent. However, we will also consider
scenarios where the optimization procedure can dynamically choose optimal termination: § 5.4, p. 103
when to return based upon inspection of the available data. practical termination: § 9.3, p. 210

1.2 the bayesian approach

Bayesian optimization does not refer to one particular algorithm but
rather to a philosophical approach to optimization grounded in Bayes-
ian inference from which an extensive family of algorithms have been
derived. Although these algorithms display significant diversity in their
details, they are bound by common themes in their design.
Optimization is fundamentally a sequence of decisions: in each it-
eration, we must choose where to make our next observation and then
whether to terminate depending on the outcome. As the outcomes of
these decisions are governed by the system under study and outside our
control, the success of optimization rests entirely on effective decision
making.
Increasing the difficulty of these decisions is that they must be made
under uncertainty, as it is impossible to know the outcome of an observa-
tion before making it. The optimization policy must therefore design each

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 5

 introduction

observation with some measure of faith that the outcome will ultimately
prove beneficial and justify the cost of obtaining it. The sequential nature
of optimization further compounds the weight of this uncertainty, as the
outcome of each observation not only has an immediate impact, but also
forms the basis on which all future decisions are made. Developing an
effective policy requires somehow addressing this uncertainty.
The Bayesian approach systematically relies on probability and Bayes-
ian inference to reason about the uncertain quantities arising during
optimization. This critically includes the objective function itself, which
is treated as a random variable to be inferred in light of our prior ex-
pectations and any available data. In Bayesian optimization, this belief
then takes an active role in decision making by guiding the optimiza-
tion policy, which may evaluate the merit of a proposed observation
location according to our belief about the value we might observe. We
introduce the key ideas of this process with examples below, starting
with a refresher on Bayesian inference.

Bayesian inference
To frame the following discussion, we offer a quick overview of Bayesian
6 The literature is vast. The following references inference as a reminder to the reader. This introduction is far from
are excellent, but no list can be complete: complete, but there are numerous excellent references available.6
d. j. c. mackay (2003). Information Theory, In- Bayesian inference is a framework for inferring uncertain features of
ference, and Learning Algorithms. Cambridge a system of interest from observations grounded in the laws of probability.
University Press. To illustrate the basic ideas, we may begin by identifying some unknown
a. o’hagan and j. forster (2004). Kendall’s feature of a given system that we wish to reason about. In the context of
Advanced Theory of Statistics. Vol. 2b: Bayes- optimization, this might represent, for example, the value of the objective
ian Inference. Arnold.
function at a given location, or the location 𝑥 ∗ or value 𝑓 ∗ of the global
j. o. berger (1985). Statistical Decision Theory optimum (1.1). We will take the first of these as a running example:
and Bayesian Analysis. Springer–Verlag. inferring about the value of an objective function at some arbitrary point
𝑥, 𝜙 = 𝑓 (𝑥). We will shortly extend this example to inference about the
entire objective function.
In the Bayesian approach to inference, all unknown quantities are
treated as random variables. This is a powerful convention as it allows us
to represent beliefs about these quantities with probability distributions
reflecting their plausible values. Inference then takes the form of an
inductive process where these beliefs are iteratively refined in light of
observed data by appealing to probabilistic identities.
As with any induction, we must start somewhere. Here we begin with
prior distribution, 𝑝 (𝜙 | 𝑥) a so-called prior distribution (or simply prior) 𝑝 (𝜙 | 𝑥), which encodes
what we consider to be plausible values for 𝜙 before observing any
7 Here we assume the location of interest 𝑥 is data.7 The prior distribution allows us to inject our knowledge about and
known, hence our conditioning the prior on experience with the system of interest into the inferential process, saving
its value.
us from having to begin “from scratch” or entertain patently absurd
possibilities. The left panel of figure 1.1 illustrates a prior distribution for
our example, indicating support over a range of values.
Once a prior has been established, the next stage of inference is
to refine our initial beliefs in light of observed data. Suppose in our

6
 1.2. the bayesian approach

prior, 𝑝 (𝜙 | 𝑥) likelihood, 𝑝 (𝑦 | 𝑥, 𝜙) posterior, 𝑝 (𝜙 | D)

𝜙 𝜙 𝜙
𝑦 𝑦

Figure 1.1: Bayesian inference for an unknown function value 𝜙 = 𝑓 (𝑥). Left: a prior distribution over 𝜙; middle: the
likelihood of the marked observation 𝑦 according to an additive Gaussian noise observation model (1.4) (prior
shown for reference); right: the posterior distribution in light of the observation and the prior (prior and
likelihood shown for reference).

example we make an observation of the objective function at 𝑥, revealing

a measurement 𝑦. In our model of optimization, the distribution of this
measurement is assumed to be determined by the value of interest 𝜙
through the observation model 𝑝 (𝑦 | 𝑥, 𝜙) (1.2). In the context of Bayesian
inference, a distribution explaining the observed values (here 𝑦) in terms
of the values of interest (here 𝜙) is known as a likelihood function or
simply a likelihood. The middle panel of figure 1.1 show the likelihood likelihood function (observation model),
– as a function of 𝜙 – for a given measurement 𝑦, here assumed to be 𝑝 (𝑦 | 𝑥, 𝜙)
generated by additive Gaussian noise (1.4).
Finally, given the observed value 𝑦, we may derive the updated poste-
rior distribution (or simply posterior) of 𝜙 by appealing to Bayes’ theorem: posterior distribution, 𝑝 (𝜙 | 𝑥, 𝑦)

𝑝 (𝜙 | 𝑥) 𝑝 (𝑦 | 𝑥, 𝜙)
𝑝 (𝜙 | 𝑥, 𝑦) = . (1.5)
𝑝 (𝑦 | 𝑥)
The posterior is proportional to the prior weighted by the likelihood of
the observed value. The denominator is a constant with respect to 𝜙 that
ensures normalization:
∫
𝑝 (𝑦 | 𝑥) = 𝑝 (𝑦 | 𝑥, 𝜙) 𝑝 (𝜙 | 𝑥) d𝜙 . (1.6)

The right panel of figure 1.1 shows the posterior resulting from the
measurement in the middle panel. The posterior represents a compromise
between our experience (encoded in the prior) and the information
contained in the data (encoded in the likelihood).
Throughout this book we will use the catchall notation D to repre- data informing posterior belief, D
sent all the information influencing a posterior belief; here the relevant
information is D = (𝑥, 𝑦), and the posterior distribution is then 𝑝 (𝜙 | D).
As mentioned previously, Bayesian inference is an inductive process
whereby we can continue to refine our beliefs through additional ob-
servation. At this point, the induction is trivial: to incorporate a new

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 7

 introduction

observation, what was our posterior serves as the prior in the context
posterior predictive, 𝑝 (𝑦 0 | 𝑥, D) of the new information, and multiplying by the likelihood and renor-
malizing yields a new posterior. We may continue in this manner as
desired.
The posterior distribution is not usually the end result of Bayesian
inference but rather a springboard enabling follow-on tasks such as
prediction or decision making, both of which are integral to Bayesian
optimization. To address the former, suppose that after deriving the
posterior (1.5), we wish to predict the result of an independent, repeated
𝑦0
noisy observation at 𝑥, 𝑦 .0 Treating the outcome as a random variable,
𝑦 we may derive its distribution by integrating our posterior belief about
𝜙 against the observation model (1.2):8
Posterior predictive distribution for a repeated ∫
measurement at 𝑥 for our running example.
The location of our first measurement 𝑦 and
𝑝 (𝑦 0 | 𝑥, D) = 𝑝 (𝑦 0 | 𝑥, 𝜙) 𝑝 (𝜙 | 𝑥, D) d𝜙; (1.7)
the posterior distribution of 𝜙 are shown for
reference. There is more uncertainty in 𝑦 0 this is known as the posterior predictive distribution for 𝑦 .0 By integrating
than 𝜙 due to the effect of observation noise. over all possible values of 𝜙 weighted by their plausibility, the posterior
predictive distribution naturally accounts for uncertainty in the unknown
8 This expression takes the same form as (1.6), objective function value; see the figure in the margin.
which is simply the (prior) predictive distribu- The Bayesian approach to decision making also relies on a posterior
tion evaluated at the actual observed value.
belief about unknown features affecting the outcomes of our decisions,
as we will discuss shortly.

Bayesian inference of the objective function

At the heart of any Bayesian optimization routine is a probabilistic belief
over the objective function. This takes the form of a stochastic process, a
stochastic process probability distribution over an infinite collection of random variables –
here the objective function value at every point. The reasoning behind
this inference is, in essence, the same as our single-point example above.
We begin by encoding any assumptions we may have about the ob-
jective function, such as smoothness or other features, in a prior process
objective function prior, 𝑝 (𝑓 ) 𝑝 (𝑓 ). Conveniently, we can specify a stochastic process via the distribu-
tion of the function values 𝝓 corresponding to an arbitrary finite set of
locations x:
𝑝 (𝝓 | x). (1.8)
chapter 2: Gaussian processes, p. 15 The family of Gaussian processes – where these finite-dimensional distri-
chapter 3: modeling with Gaussian processes,
butions are multivariate Gaussian – is especially convenient and widely
p. 45 used in Bayesian optimization. We will explore this model class in depth
in the following three chapters; here we provide a motivating illustration.
chapter 4: model assessment, selection, and
averaging, p. 67
Figure 1.2 shows a Gaussian process prior on a one-dimensional
objective function, constructed to reflect a minimal set of assumptions
we will elaborate on later in the book:
differentiability: § 2.6, p. 30 • that the objective function is smooth (that is, infinitely differentiable),
characteristic length scales: § 3.4, p. 56 • that correlations among function values have a characteristic scale, and
stationarity: § 3.2, p. 50 • that the function’s expected behavior does not depend on location (that
is, the prior process is stationary).

8
 1.2. the bayesian approach

prior mean prior 95% credible interval samples

Figure 1.2: An example prior process for an objective defined on an interval. We illustrate the marginal belief at every
location with its mean and a 95% credible interval and also show three example functions sampled from the
prior process.

We summarize the marginal belief of the model, for each point in the
domain showing the prior mean (dark blue) and 95% credible interval
(light blue) for the corresponding function value. We also show three
functions sampled from the prior process, each exhibiting the assumed
behavior. We encourage the reader to become comfortable with this
plotting convention, as we will use it throughout this book. In particular
we eschew axis labels, as they are always the same: the horizontal axis plotting conventions
represents the domain X and the vertical axis the function value. Further,
we do not mark units on axes to stress relative rather than absolute
behavior, as scale is arbitrary in this illustration.
We can encode a vast array of information into the prior process
and can model significantly more complex structure than in this sim-
ple example. We will explore the world of possibilities in chapter 3,
including interaction at different scales, nonstationarity, low intrinsic nonstationarity, warping: § 3.4, p. 56
dimensionality, and more. low intrinsic dimensionality: § 3.5, p. 61
With the prior process in hand, suppose we now make a set of
observations at some locations x, revealing corresponding values y; we
aggregate this information into a dataset D = (x, y). Bayesian inference observed data, D = (x, y)
accounts for these observations by forming the posterior process 𝑝 (𝑓 | D). objective function posterior, 𝑝 (𝑓 | D)
The derivation of the posterior process can be understood as a two-
stage process. First we consider the impact of the data on the correspond-
ing function values 𝝓 alone (1.5):
9 The given expression sweeps some details un-
𝑝 (𝝓 | D) ∝ 𝑝 (𝝓 | x) 𝑝 (y | x, 𝝓). (1.9) der the rug. A careful derivation of the pos-
terior process proceeds by finding the poste-
The quantities on the right-hand side are known: the first term is given rior of an arbitrary finite-dimensional vector
by the prior process (1.8), and the second by the observation model (1.3), 𝝓∗ = 𝑓 (x∗ ):
which serves the role of a likelihood. We now extend the posterior on 𝝓
to all of 𝑓 :9 ∫
𝑝 (𝝓 ∗ | x∗ , D) =
∫
𝑝 (𝑓 | D) = 𝑝 (𝑓 | x, 𝝓) 𝑝 (𝝓 | D) d𝝓. (1.10) 𝑝 (𝝓∗ | x∗ , x, 𝝓) 𝑝 (𝝓 | D) d𝝓,

which specifies the process. The distributions

The posterior encapsulates our belief regarding the objective in light of
on the right-hand side are known: the pos-
the data, incorporating both the assumptions of the prior process and terior on 𝝓 is in (1.9), and the posterior on
the information contained in the observations. 𝝓∗ given the exact function values 𝝓 can be
We illustrate an example posterior in figure 1.3, where we have con- found by computing their joint prior (1.8) and
conditioning.
ditioned our prior from figure 1.2 on three exact observations. As the

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 9

introduction

observations posterior mean posterior 95% credible interval samples

Figure 1.3: The posterior process for our example scenario in figure 2.1 conditioned on three exact observations.

acquisition function next observation location

Figure 1.4: A prototypical acquisition function corresponding to our example posterior from figure 1.3.

observations are assumed to be exact, the objective function posterior

collapses onto the observed values. The posterior mean interpolates
through the data, and the posterior credible intervals reflect increased
certainty regarding the function near the observed locations. Further,
the posterior continues to reflect the structural assumptions encoded
in the prior, demonstrated by comparing the behavior of the samples
drawn from the posterior process to those drawn from the prior.

Uncertainty-aware optimization policies

Bayesian inference provides an elegant means of reasoning about an
uncertain objective function, but the success of optimization is measured
not by the fidelity of our beliefs but by the outcomes of our actions.
These actions are determined by the optimization policy, which exam-
ines available data to design each successive observation location. Each
of these decisions is fraught with uncertainty, as we must commit to each
observation before knowing its result, which will form the context of all
following decisions. Bayesian inference enables us to express this uncer-
tainty, but effective decision making additionally requires us to establish
preferences over outcomes and act to maximize those preferences.
chapter 5: decision theory for optimization, To proceed we need to establish a framework for decision making
p. 87 under uncertainty, an expansive subject with a world of possibilities.
chapter 6: utility functions for optimization, A natural and common choice is Bayesian decision theory, the subject
p. 109 of chapters 5–6. We will discuss this and other approaches to policy
chapter 7: common Bayesian optimization
construction at length in chapter 7 and derive popular optimization
policies, p. 123 policies from first principles.
Ignoring details in policy design, a thread running through all Bayes-
ian optimization policies is a uniform handling of uncertainty in the
objective function and the outcomes of observations via Bayesian infer-

10
 1.2. the bayesian approach

ence. Instrumental in connecting our beliefs about the objective function

to decision making is the posterior predictive distribution (1.7), represent-
ing our belief about the outcomes of proposed observations. Bayesian
optimization policies are designed with reference to this distribution,
which guides the policy in discriminating between potential actions.
In practice, Bayesian optimization policies are defined indirectly by
optimizing a so-called acquisition function assigning a score to potential acquisition functions: § 5, p. 88
observation locations commensurate with their perceived ability to ben-
efit the optimization process. Acquisition functions tend to be cheap to
evaluate with analytically tractable gradients, allowing the use of off-
the-shelf optimizers to efficiently design each observation. Numerous
acquisition functions have been proposed for Bayesian optimization, each
derived from different considerations. However, all notable acquisition
functions address the classic tension between exploitation – sampling
where the objective function is expected to be high – and exploration –
sampling where we are uncertain about the objective function to inform
future decisions. These opposing concerns must be carefully balanced
for effective global optimization.
An example acquisition function is shown in figure 1.4, correspond- example and discussion
ing to the posterior from figure 1.3. Consideration of the exploitation–
exploration tradeoff is apparent: this example acquisition function attains
relatively large values both near local maxima of the posterior mean
and in regions with significant marginal uncertainty. Local maxima of
the acquisition function represent optimal compromises between these
concerns. Note that the acquisition function vanishes at the location of
the current observations: the objective function values at these locations
are already known, so observing there would be pointless. Maximizing
the acquisition function determines the policy; here the policy chooses
to search around the local optimum on the right-hand side.
Figure 1.5 demonstrates an entire session of Bayesian optimization,
beginning from the belief and initial decision from figure 1.4 and pro-
gressing iteratively following algorithm 1.1. The red function shows the
true (unknown) objective function, whose maximum is near the center of
the domain. The running marks below each posterior show the locations
of each measurement made, progressing in sequence from top to bottom,
and we show the objective function posterior at four waypoints.
Dynamic consideration of the exploitation–exploration tradeoff is
evident in the algorithm’s behavior. The first two observations map out
the neighborhood of the initially best-seen point, exhibiting exploitation.
Once sufficiently explored, the policy continues exploitation around the
second best-seen point, discovering and refining the global optimum in
iterations 7–8. Finally, the policy switches to exploration in iterations
13–19, systematically covering the domain to ensure nothing has been
missed. At termination, there is clear bias in the collected data toward
higher objective values, and all remaining uncertainty is in regions where
the credible intervals indicate the optimum is unlikely to reside.
The “magic” of Bayesian optimization is that the intuitive behavior
of this optimization policy is not the result of ad hoc design, but rather

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 11

introduction

objective function

1
2

exploitation

3
4
5
6
7
8
9
10
11
12

exploitation

13
14
15
16
17
18
19

exploration

Figure 1.5: The posterior after the indicated number of steps of an example Bayesian optimization policy, starting from
the posterior in figure 1.4. The marks show the points chosen by the policy, progressing from top to bottom.
Observations sufficiently close to the optimum are marked in red; the optimum was located on iteration 7.

12
 1.2. the bayesian approach

emerges automatically through the machinery of Gaussian processes and

Bayesian decision theory that we will develop over the coming chapters.
In this framework, building an optimization policy boils down to:
• choosing a model of the objective function,
• deciding what sort of data we seek to obtain, and
• systematically transforming these beliefs and preferences into an opti-
mization policy.
Over the following chapters, we will develop tools for achieving each
of these goals: Gaussian processes (chapters 2–4) for expressing what
we believe about the objective function, utility functions (chapter 6)
for expressing what we value in data, and Bayesian decision theory
(chapter 5) for building optimization policies aware of the uncertainty
encoded in the model and guided by the preferences encoded in the utility
function. In chapter 7 we will combine these fundamental components
to realize complete Bayesian optimization policies, at which point we
will be equipped to replicate this example from first principles.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 13

 2
GAUSSIAN PROCESSES

The central object in optimization is an objective function 𝑓 : X → ℝ, and

the primary challenge in algorithm design is inherent uncertainty about
this function: most importantly, where is the function maximized and
what is its maximal value? Prior to optimization, we may very well have
no idea. Optimization affords us the opportunity to acquire information
about the objective – through observations of our own design – to shed
light on these questions. However, this process is itself fraught with
uncertainty, as we cannot know the outcomes and implications of these
observations at the time of their design. Notably, we face this uncertainty
even when we have a closed-form expression for the objective function,
a favorable position as many objectives act as “black boxes.”
Reflecting on this situation, diaconis posed an intriguing question:1 1 p. diaconis (1988). Bayesian Numerical Anal-
“what does it mean to ‘know’ a function?” The answer is unclear when an ysis. In: Statistical Decision Theory and Related
Topics iv.
analytic expression, which might at first glance seem to encapsulate the
essence of the function, is insufficient to determine features of interest.
However, diaconis argued that although we may not know everything
about a function, we often have some prior knowledge that can facilitate
a numerical procedure such as optimization. For example, we may expect
an objective function to be smooth (or rough), or to assume values in
a given range, or to feature a relatively simple underlying trend, or to
depend on some hidden low-dimensional representation we hope to
uncover.2 All of this knowledge could be instrumental in accelerating 2 We will explore all of these possibilities in the
optimization if it could be systematically captured and exploited. next chapter, p. 45.
Having identifiable information about an objective function prior to
optimization motivates the Bayesian approach we will explore through-
out this book. We will address uncertainty in the objective function
through the unifying framework of Bayesian inference, treating 𝑓 – as Bayesian inference of the objective function:
well as ancillary quantities such as 𝑥 ∗ and 𝑓 ∗ (1.1) – as random variables § 1.2, p. 8
to be inferred from observations revealed during optimization.
To pursue this approach, we must first determine how to build useful
prior distributions for objective functions and how to compute a poste-
rior belief given observations. If the system under investigation is well
understood, we may be able to identify an appropriate parametric form
𝑓 (𝑥; 𝜽 ) and infer the parameters 𝜽 directly. This approach is likely the
best course of action when possible;3 however, many objective functions 3 v. dalibard et al. (2017). boat: Building Auto-
have no obvious parametric form, and most models used in Bayesian Tuners with Structured Bayesian Optimization.
www 2017.
optimization are thus nonparametric to avoid undue assumptions.4
In this chapter we will introduce Gaussian processes (gps), a conve- 4 The term“nonparametric” is something of a
misnomer. A nonparametric objective func-
nient class of nonparametric regression models widely used in Bayesian tion model has parameters but their dimen-
optimization. We will begin by defining Gaussian processes and deriving sion is infinite – we effectively parametrize
some basic properties, then demonstrate how to perform inference from the objective by its value at every point.
observations. In the case of exact observation and additive Gaussian
noise, we can perform this inference exactly, resulting in an updated
posterior Gaussian process. We will continue by considering some the-
oretical properties of Gaussian processes relevant to optimization and
inference with non-Gaussian observation models.

This material will be published by Cambridge University Press as Bayesian Optimization. 15

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 gaussian processes

The literature on Gaussian processes is vast, and we do not intend this

chapter to serve as a standalone introduction but rather as companion to
the existing literature. Although our discussion will be comprehensive,
our focus on optimization will sometimes bias its scope. For a broad
5 c. e. rasmussen and c. k. i. williams (2006). overview, the interested reader may consult rasmussen and williams’s
Gaussian Processes for Machine Learning. mit classic monograph.5
Press.

2.1 definition and basic properties

multivariate normal distribution: appendix a, A Gaussian process is an extension of the familiar multivariate normal dis-
p. 295 tribution suitable for modeling functions on infinite domains. Gaussian
processes inherit the convenient mathematical properties of the multi-
chapter 3: modeling with Gaussian processes, variate normal distribution without sacrificing computational tractability.
p. 45 Further, by modifying the structure of a gp, we can model functions with
a rich variety of behavior; we will explore this capability in the next
6 p. hennig et al. (2015). Probabilistic numerics chapter. This combination of mathematical elegance and flexibility in
and uncertainty in computations. Proceedings modeling has established Gaussian processes as the workhorse of Bayes-
ian approaches to numerical tasks, including optimization.6, 7
of the Royal Society A: Mathematical, Physical
and Engineering Sciences 471(2179):20150142.
7 p. hennig et al. (2022). Probabilistic Numerics:
Computation as Machine Learning. Cambridge Definition
University Press.
Consider an objective function 𝑓 : X → ℝ of interest over an arbitrary
8 If X is finite, there is no distinction between infinite domain X.8 We will take a nonparametric approach and reason
a Gaussian process and a multivariate normal about the function as an infinite collection of random variables, one
distribution, so only the infinite case is inter-
esting for this discussion. corresponding to the function value at every point in the domain. Mutual
dependence between these random variables will then determine the
statistical properties of the function’s shape.
It is perhaps not immediately clear how we can specify a useful distri-
bution over infinitely many random variables, a construction known as a
stochastic process. However, a result known as the Kolmogorov extension
theorem allows us to construct a stochastic process by defining only the
9 b. øksendal (2013). Stochastic Differential distribution of arbitrary finite sets of function values, subject to natural
Equations: An Introduction with Applications. consistency constraints.9 For a Gaussian process, these finite-dimensional
Springer–Verlag. [§ 2.1]
distributions are all multivariate Gaussian, hence its name.
In this light, we build a Gaussian process by replacing the parameters
10 Writing the process as if it were a function- in the finite-dimensional case – a mean vector and a positive semidefinite
valued probability density function is an abuse covariance matrix – by analogous functions over the domain. We specify
of notation, but a useful and harmless one.
a Gaussian process on 𝑓:10

𝑝 (𝑓 ) = GP (𝑓 ; 𝜇, 𝐾)

mean function, 𝜇 by a mean function 𝜇 : X → ℝ and a positive semidefinite covariance

covariance function (kernel), 𝐾 function (or kernel) 𝐾 : X × X → ℝ. The mean function determines the
value of objective at 𝑥, 𝜙 expected function value 𝜙 = 𝑓 (𝑥) at any location 𝑥:

𝜇 (𝑥) = 𝔼[𝜙 | 𝑥],

thus serving as a location parameter representing the function’s central

tendency. The covariance function determines how deviations from the

16
 2.1. definition and basic properties

mean are structured, encoding expected properties of the function’s

behavior. Defining 𝜙 0 = 𝑓 (𝑥 0), we have: value of objective at 𝑥 ,0 𝜙 0

𝐾 (𝑥, 𝑥 0) = cov[𝜙, 𝜙 0 | 𝑥, 𝑥 0]. (2.1)

The mean and covariance functions of the process allow us to com-

pute any finite-dimensional marginal distribution on demand. Let x ⊂ X
be finite and let 𝝓 = 𝑓 (x) be the corresponding function values, a vector- values of objective at x, 𝝓 = 𝑓 (x)
valued random variable. For the Gaussian process (2.1), the distribution
of 𝝓 is multivariate normal with parameters determined by the mean
and covariance functions:

𝑝 (𝝓 | x) = N (𝝓; 𝝁, 𝚺), (2.2)

where
𝝁 = 𝔼[𝝓 | x] = 𝜇 (x); 𝚺 = cov[𝝓 | x] = 𝐾 (x, x). (2.3)

Here 𝐾 (x, x) is the matrix formed by evaluating the covariance function

for each pair of points: Σ𝑖 𝑗 = 𝐾 (𝑥𝑖 , 𝑥 𝑗 ), also called the Gram matrix of x. Gram matrix of x, 𝚺 = 𝐾 (x, x)
In many ways, Gaussian processes behave like “really big” Gaussian
distributions, and one can intuit many of their properties from this
heuristic alone. For example, the Gaussian marginal property in (2.2–
2.3) corresponds precisely with the analogous formula in the finite- 11 In fact, this is precisely the consistency re-
dimensional case (a.13). Further, this property automatically ensures quired by the Kolmogorov extension theorem
mentioned on the facing page.
global consistency in the following sense.11 If x is an arbitrary set of
points and x0 ⊃ x is a superset, then we arrive at the same belief about
𝝓 whether we compute it directly from (2.2–2.3) or indirectly by first marginalizing multivariate normal
computing 𝑝 (𝝓 0 | x0) then marginalizing x0 \ x (a.13). distributions, § a.2, p. 299

Example and basic properties

Let us construct and explore an explicit Gaussian process for a function
on the interval X = [0, 30]. For the mean function we take the zero
function 𝜇 ≡ 0, indicating a constant central tendency. For the covariance
function, we take the prototypical squared exponential covariance: squared exponential covariance: § 3.3, p. 51


𝐾 (𝑥, 𝑥 0) = exp − 12 |𝑥 − 𝑥 0 | 2 . (2.4)

Let us pause to consider the implications of this choice. First, note that
var[𝜙 | 𝑥] = 𝐾 (𝑥, 𝑥) = 1 at every point 𝑥 ∈ X, and thus the covariance 𝐾 (𝑥, 𝑥 0 )
function (2.4) also measures the correlation between the function values 1
𝜙 and 𝜙 .0 This correlation decreases with the distance between 𝑥 and 𝑥 ,0
falling from unity to zero as these points become increasingly separated;
see the illustration in the margin. We can loosely interpret this as a
statistical consequence of continuity: function values at nearby locations
|𝑥 − 𝑥 0 |
are highly correlated, whereas function values at distant locations are 0
effectively independent. This assumption also implies that observing 0 3
the function at some point 𝑥 provides nontrivial information about the The squared exponential covariance (2.4) as a
function at sufficiently nearby locations (roughly when |𝑥 − 𝑥 0 | < 3). We function of the distance between inputs.
will explore this implication further shortly.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 17

gaussian processes

prior mean prior 95% credible interval samples

Figure 2.1: Our example Gaussian process on the domain X = [0, 30]. We illustrate the marginal belief at every location
with its mean and a 95% credible interval and also show three example functions sampled from the process.

predictive credible intervals For a Gaussian process, the marginal distribution of any single func-
tion value is univariate normal (2.2):

𝑝 (𝜙 | 𝑥) = N (𝜙; 𝜇, 𝜎 2 ); 𝜇 = 𝜇 (𝑥); 𝜎 2 = 𝐾 (𝑥, 𝑥), (2.5)

where we have abused notation slightly by overloading the symbol 𝜇.

This allows us to derive pointwise credible intervals; for example, the
familiar 𝜇 ± 1.96𝜎 is a 95% credible interval for 𝜙. Examining our example
gp, the marginal distribution of every function value is in fact standard
normal. We provide a rough visual summary of the process via its mean
function and pointwise 95% predictive credible intervals in figure 2.1.
There is nothing terribly exciting we can glean from these marginal
distributions alone, and no interesting structure in the process is yet
apparent.

Sampling
We may gain more insight by inspecting samples drawn from our exam-
ple process reflecting the joint distribution of function values. Although
it is impossible to represent an arbitrary function on X in finite memory,
we can approximate the sampling process by taking a dense grid x ⊂ X
sampling from a multivariate normal and sampling the corresponding function values from their joint multi-
distribution: § a.2, p. 299 variate normal distribution (2.2). Plotting the sampled vectors against
the chosen grid reveals curves approximating draws from the Gaussian
process. Figure 2.1 illustrates this procedure for our example using a grid
of 1000 equally spaced points. Each sample is smooth and has several
local optima distributed throughout the domain – for some applications,
this might be a reasonable model for an objective function on X.

2.2 inference with exact and noisy observations

We now turn to our attention to inference: given a Gaussian process
prior on an objective function, how can we condition this initial belief
on observations obtained during optimization?
example and discussion Let us look at an example to build intuition before diving into the
details. Figure 2.2 shows the effect of conditioning our example gp from

18
 2.2. inference with exact and noisy observations

observations posterior mean posterior 95% credible interval samples

Figure 2.2: The posterior for our example scenario in figure 2.1 conditioned on three exact observations.

the previous section on three exact measurements of the function. The

updated belief reflects both our prior assumptions and the information
contained in the data, the hallmark of Bayesian inference. To elaborate,
the posterior mean smoothly interpolates through the observed values,
agreeing with both the measured values and the smoothness encoded
in the prior covariance function. The posterior credible intervals are
reduced in the neighborhood of the measured locations – where the
prior covariance function encodes nontrivial dependence on at least one
observed value – and vanish where the function value has been exactly
determined. On the other hand, our marginal belief remains effectively
unchanged from the prior in regions sufficiently isolated from the data,
where the prior covariance function encodes effectively no correlation.
Conveniently, inference is straightforward for the pervasive observa-
tion models of exact measurement and additive Gaussian noise, where
the self-conjugacy of the normal distribution yields a Gaussian process
posterior with updated parameters we can compute in closed form. The
reasoning underlying inference for both observation models is identical
and is subsumed by a flexible general argument we will present first.

Inference from arbitrary jointly Gaussian observations

We may exactly condition a Gaussian process 𝑝 (𝑓 ) = GP (𝑓 ; 𝜇, 𝐾) on the
observation of any vector y sharing a joint Gaussian distribution with 𝑓 : vector of observed values, y

     !
𝑓 𝜇 𝐾 𝜅>
𝑝 (𝑓, y) = GP ; , . (2.6)
y m 𝜅 C

This notation, analogous to (a.12), extends the Gaussian process on 𝑓 to 12 We assume C is positive definite; if it were only
include the entries of y; that is, we assume the distribution of any finite positive semidefinite, there would be wasteful
linear dependence among observations.
subset of function and/or observed values is multivariate normal. We
specify the joint distribution via the marginal distribution of y:12 observation mean and covariance, m, C

𝑝 (y) = N (y; m, C) (2.7)

and the cross-covariance function between y and 𝑓 : cross-covariance between observations and
function values, 𝜅
𝜅 (𝑥) = cov[y, 𝜙 | 𝑥]. (2.8)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 19

 gaussian processes

Although it may seem absurd that we could identify and observe a vector
satisfying such strong restrictions on its distribution, we can already
deduce several examples from first principles, including:
inference from exact observations: § 2.2, p. 22 • any vector of function values (2.2),
affine transformations: § a.2, p. 298 • any affine transformation of function values (a.10), and
derivatives and expectations: § 2.6, p. 30 • limits of such quantities, such as partial derivatives or expectations.
Further, we may condition on any of the above even if corrupted by
independent additive Gaussian noise, as we will shortly demonstrate.
conditioning a multivariate normal We may condition the joint distribution (2.6) on y analogously to the
distribution: § a.2, p. 299 finite-dimensional case (a.14), resulting in a Gaussian process posterior
on 𝑓. Writing D = y for the observed data, we have:

𝑝 (𝑓 | D) = GP (𝑓 ; 𝜇D , 𝐾D ), (2.9)
posterior mean and covariance, 𝜇 D , 𝐾D where
𝜇D (𝑥) = 𝜇 (𝑥) + 𝜅 (𝑥)> C−1 (y − m);
(2.10)
𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − 𝜅 (𝑥)> C−1 𝜅 (𝑥 0).
13 This is a useful exercise! The result will be a This can be verified by computing the joint distribution of an arbitrary
stochastic process with multivariate normal finite set of function values and y and conditioning on the latter (a.14).13
finite-dimensional distributions, a Gaussian
process by definition (2.5).
The above result provides a simple procedure for gp posterior infer-
ence from any vector of observations satisfying (2.6):
1. compute the marginal distribution of y (2.7),
2. derive the cross-covariance function 𝜅 (2.8), and
3. find the posterior distribution of 𝑓 via (2.9–2.10).
We will realize this procedure for several special cases below. However,
we will first demonstrate how we may seamlessly handle measurements
corrupted by additive Gaussian noise and build intuition for the posterior
distribution by dissecting its moments in terms of the statistics of the
observations and the correlation structure of the prior.

Corruption by additive Gaussian noise

We pause to make one observation of immense practical importance:
any vector satisfying (2.6) would continue to suffice even if corrupted
by independent additive Gaussian noise, and thus we can use the above
result to condition a Gaussian process on noisy observations as well.
Suppose that rather than observing y exactly, our measurement
noisy observation of y, z mechanism only allowed observing z = y + 𝜺 instead, where 𝜺 is a vector
vector of random errors, 𝜺 of random errors independent of y. If the errors are normally distributed
noise covariance matrix, N with mean zero and known (arbitrary) covariance N:

𝑝 (𝜺 | N) = N (𝜺; 0, N), (2.11)

sums of normal vectors: § a.2, p. 300 then we have
𝑝 (z | N) = N (z; m, C + N); cov[z, 𝜙 | 𝑥] = cov[y, 𝜙 | 𝑥] = 𝜅 (𝑥).

20
 2.2. inference with exact and noisy observations

Thus we can condition on an observation of the corrupted vector z by

simply replacing C with C + N in the prior (2.6) and posterior (2.10).14 14 Assuming zero-mean errors is not strictly nec-
Note that the posterior converges to that from a direct observation of y essary but is overwhelmingly common in prac-
tice. A nonzero mean 𝔼[𝜺 ] = n is possible by

if we take the noise covariance N → 0 in the positive semidefinite cone, further replacing (y − 𝝁) with y − [𝝁 + n]
a reassuring result. in (2.10), where 𝝁 + n = 𝔼[y].

Interpretation of posterior moments

The moments of the posterior Gaussian process (2.10) contain update
terms adjusting the prior moments in light of the data. These updates
have intuitive interpretations in terms of the nature of the prior process
and the observed values, which we may unravel with some care.
We can gain some initial insight by considering the case where we
observe a single value with 𝑦 distribution N (𝑦; 𝑚, 𝑠 2 ) and breaking down
its impact on our belief. Consider an arbitrary function value 𝜙 with
prior distribution N (𝜙; 𝜇, 𝜎 2 ) (2.5) and define 𝑧-score of measurement 𝑦, 𝑧
𝑦 −𝑚
𝑧=
𝑠
to be the 𝑧-score of the observed value 𝑦 and correlation between measurement 𝑦 and
function value 𝜙, 𝜌
𝜅 (𝑥)
𝜌 = corr[𝑦, 𝜙 | 𝑥] =
𝜎𝑠
to be the correlation between 𝑦 and 𝜙. Then the posterior mean and posterior moments of 𝜙 from a scalar
standard deviation of 𝜙 are, respectively: observation
√︁
𝜇 + 𝜎𝜌𝑧; 𝜎 1 − 𝜌2. (2.12)
The 𝑧-score of the posterior mean, with respect to the prior distri- interpretation of moments
bution of 𝜙, is 𝜌𝑧. An independent measurement with 𝜌 = 0 thus leaves
the prior mean unchanged, whereas a perfectly dependent measurement
with |𝜌 | = 1 shifts the mean up or down by 𝑧 standard deviations (de-
pending on the sign of the correlation) to match the magnitude of the
measurement’s 𝑧-score. Measurements with partial dependence result in
outcomes between these extremes. Further, surprising measurements – √︁
𝜎 1 − 𝜌2
that is, those with large |𝑧| – yield larger shifts in the mean, whereas an
𝜎
entirely expected measurement with 𝑦 = 𝑚 leaves the mean unchanged.
Turning to the posterior standard deviation, the measurement re-
duces our uncertainty in 𝜙 by a factor depending on the correlation
𝜌, but not on the value observed. An independent measurement again
leaves the prior intact, whereas a perfectly dependent measurement col-
lapses the posterior standard deviation to zero as the value of 𝜙 would be 0
|𝜌 |
completely determined. The relative reduction in posterior uncertainty 0 1
as a function of the absolute correlation is illustrated in the margin.
The posterior standard deviation of 𝜙 as a
In the case of vector-valued observations, we can interpret similar function of the strength of relationship with
structure in the posterior, although dependence between entries of y 𝑦, |𝜌 |.
must also be accounted for. We may factor the observation covariance
matrix as
C = SPS, (2.13)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 21

 gaussian processes

√
where S is diagonal with 𝑆𝑖𝑖 = 𝐶𝑖𝑖 = std[𝑦𝑖 ] and P = corr[y] is the
observation correlation matrix. We may then rewrite the posterior mean
of 𝜙 as
𝜇 + 𝜎𝝆> P−1 z,
where z and 𝝆 represent the vectors of measurement 𝑧-scores and the
cross-correlation between 𝜙 and y, respectively:

𝑦𝑖 − 𝑚𝑖 [𝜅 (𝑥)] 𝑖
𝑧𝑖 = ; 𝜌𝑖 = .
𝑠𝑖 𝜎𝑠𝑖
15 It can be instructive to contrast the behavior The posterior mean is now in the same form as the scalar case (2.12),
of the posterior when conditioning on two with the introduction of the observation correlation matrix moderating
highly correlated values versus two indepen-
dent ones. In the former case, the posterior
the 𝑧-scores to account for dependence between the observed values.15
does not change much as a result of the sec- The posterior standard deviation of 𝜙 in the vector-valued case is
ond measurement, as dependence reduces the √︃
effective number of measurements.
𝜎 1 − 𝝆> P−1𝝆,

again analogous to (2.12). Noting that the inverse correlation matrix

16 P is congruent to C (2.13) and is thus positive P−1 is positive definite,16 the posterior covariance again reflects a global
definite from Sylvester’s law of inertia. reduction in the marginal uncertainty of every function value. In fact, the
joint distribution of any set of function values has reduced uncertainty
17 For positive semidefinite A, B, |A| ≤ |A + B|. in the posterior in terms of the differential entropy (a.16), as17

|𝐾 (x, x) − 𝜅 (x)> C−1 𝜅 (x)| ≤ |𝐾 (x, x)|.

The reduction of uncertainty again depends on the strength of depen-

dence between function values and the observed data, with independence
(𝝆 = 0) resulting in no change. The reduction also depends on the pre-
18 The Löwner order is the partial order induced cision of the measurements: all other things held equal, observations
by the convex cone of positive-semidefinite with greater precision in terms of the Löwner order18 on the precision
matrices. For symmetric A, B, we define A ≺ B
if and only if B − A is positive definite:
matrix C−1 provide a globally better informed posterior. In particular, as
(C + N) −1 ≺ C−1 for any noise covariance N, noisy measurements (2.11)
k. löwner (1934). Über monotone Matrixfunk- categorically provide less information about the function than direct
tionen. Mathematische Zeitschrift 38:177–216.
observations, as one should hope.

Inference with exact function evaluations

We will now explicitly demonstrate the general process of Gaussian
process inference for important special cases, beginning with the simplest
possible observation mechanism: exact observation.
Suppose we have observed 𝑓 at some set of locations x, revealing the
observed data, D = (x, 𝝓) corresponding function values 𝝓 = 𝑓 (x), and let D = (x, 𝝓) denote this
dataset. The observed vector shares a joint Gaussian distribution with
any other set of function values by the gp assumption on 𝑓 (2.2), so we
may follow the above procedure to derive the posterior. The marginal
distribution of 𝝓 is Gaussian (2.3):

𝑝 (𝝓 | x) = N (𝝓; 𝝁, 𝚺),

22
 2.2. inference with exact and noisy observations

and the cross-covariance between an arbitrary function value and 𝝓 is

by definition given by the covariance function:

𝜅 (𝑥) = cov[𝝓, 𝜙 | x, 𝑥] = 𝐾 (x, 𝑥).

Appealing to (2.9–2.10) we have:

𝑝 (𝑓 | D) = GP (𝑓 ; 𝜇D , 𝐾D ),
where
𝜇D (𝑥) = 𝜇 (𝑥) + 𝐾 (𝑥, x)𝚺−1 (𝝓 − 𝝁);
(2.14)
𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − 𝐾 (𝑥, x)𝚺−1 𝐾 (x, 𝑥 0).

Our previous figure 2.2 illustrates the posterior resulting from con- example and discussion
ditioning our gp prior in figure 2.1 on three exact measurements, with
high-level analysis of its behavior in the accompanying text.

Inference with function evaluations corrupted by additive Gaussian noise

With the notable exception of optimizing the output of a deterministic
computer program or simulation, observations of an objective function
are typically corrupted by noise due to measurement limitations or statis-
tical approximation; we must be able to handle such noisy observations to
maximize utility. Fortunately, in the important case of additive Gaussian
noise, we may perform exact inference following the general procedure
described above. In fact, the derivation below follows directly from our
previous discussion on arbitrary additive Gaussian noise, but the case of
additive Gaussian noise in function evaluations is important enough to arbitrary additive Gaussian noise: § 2.2, p. 20
merit its own discussion.
Suppose we make observations of 𝑓 at locations x, revealing cor-
rupted values y = 𝝓 + 𝜺. Suppose the measurement errors 𝜺 are indepen- 19 Allowing nondiagonal N departs from our typ-
dent of 𝝓 and normally distributed with mean zero and covariance N, ical convention of assuming conditional inde-
pendence between observations (1.3), but do-
which may optionally depend on x:
ing so does not complicate inference, so there
is no harm in this generality.
𝑝 (𝜺 | x, N) = N (𝜺; 0, N). (2.15)

As before we aggregate the observations into a dataset D = (x, y).

The observation noise covariance can in principle be arbitrary;19 special case: independent homoskedastic
however, the most common models in practice are independent ho- noise
moskedastic noise with scale 𝜎𝑛 :

N = 𝜎𝑛2 I, (2.16)

and independent heteroskedastic noise with scale depending on location special case: independent heteroskedastic
according to a function 𝜎𝑛 : X → ℝ ≥0 : noise

N = diag 𝜎𝑛2 (x). (2.17)

For a given observation location 𝑥, we will simply write 𝜎𝑛 for the observation noise scale, 𝜎𝑛
associated noise scale, leaving any dependence on 𝑥 implicit.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 23

gaussian processes

observations posterior mean posterior 95% credible interval

Figure 2.3: Posteriors for our example gp from figure 2.1 conditioned on 15 noisy observations with independent ho-
moskedastic noise (2.16). The signal-to-noise ratio is 10 for the top example, 3 for the middle example, and 1 for
the bottom example.

The prior distribution of the observations is now multivariate normal

(2.3, a.15):
𝑝 (y | x, N) = N (y; 𝝁, 𝚺 + N). (2.18)
Due to independence of the noise, the cross-covariance remains the same
as in the exact observation case:

𝜅 (𝑥) = cov[y, 𝜙 | x, 𝑥] = 𝐾 (x, 𝑥).

Conditioning on the observed value now yields a gp posterior with

𝜇D (𝑥) = 𝜇 (𝑥) + 𝐾 (𝑥, x) (𝚺 + N) −1 (y − 𝝁);

(2.19)
𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − 𝐾 (𝑥, x) (𝚺 + N) −1 𝐾 (x, 𝑥 0).

homoskedastic example and discussion Figure 2.3 shows a sequence of posterior distributions resulting from
conditioning our example gp on data corrupted by increasing levels of
homoskedastic noise (2.16). As the noise level increases, the observations
have diminishing influence on our belief, with some extreme values
eventually being partially explained away as outliers. As measurements
are assumed to be inexact, the posterior mean is not compelled to inter-
polate perfectly through the observations, as in the exact case (figure

24
 2.2. inference with exact and noisy observations

observations posterior 95% credible interval, 𝑦

posterior mean posterior 95% credible interval, 𝜙

Figure 2.4: The posterior distribution for our example gp from figure 2.1 conditioned on 50 observations with heteroskedas-
tic observation noise (2.17). We show predictive credible intervals for both the latent objective function and
noisy observations; the standard deviation of the observation noise increases linearly from left to right.

2.2). Further, with increasing levels of noise, our posterior belief reflects
significant residual uncertainty in the function, even in regions with
multiple nearby observations.
We illustrate an example of Gaussian process inference with het- heteroskedastic example and discussion
eroskedastic noise (2.17) in figure 2.4, where the signal-to-noise ratio
decreases smoothly from left-to-right over the domain. Although the
observations provide relatively even coverage, our posterior uncertainty
is minimal on the left-hand side of the domain – where the measure-
ments provide maximal information – and increases as our observations
become more noisy and less informative.
We will often require the posterior predictive distribution for a noisy posterior predictive distribution for noisy
measurement 𝑦 that would result from observing at a given location 𝑥. observations
The posterior distribution on 𝑓 (2.19) provides the posterior predictive
distribution for the latent function value 𝜙 = 𝑓 (𝑥) (2.5):

𝑝 (𝜙 | 𝑥, D) = N (𝜙; 𝜇, 𝜎 2 ); 𝜇 = 𝜇D (𝑥); 𝜎 2 = 𝐾D (𝑥, 𝑥),

but does not account for the effect of observation noise. In the case of
independent additive Gaussian noise (2.16–2.17), deriving the posterior
predictive distribution is trivial; we have (a.15):

𝑝 (𝑦 | 𝑥, D, 𝜎𝑛 ) = N (𝑦; 𝜇, 𝜎 2 + 𝜎𝑛2 ). (2.20)

This predictive distribution is illustrated in figure 2.4; the credible inter-

vals for noisy measurements reflect inflation of the credible intervals for
the underlying function value commensurate with the scale of the noise.
If the noise contains nondiagonal correlation structure, we must predictive distribution with correlated noise
account for dependence between training and test errors in the predictive
distribution. The easiest way to proceed is to recognize that the noisy
observation process 𝑦 = 𝜙 + 𝜀, as a function of 𝑥, is itself a Gaussian covariance function for noisy measurements,
process with mean function 𝜇 and covariance function 𝐶

𝐶 (𝑥, 𝑥 0) = cov[𝑦, 𝑦 0 | 𝑥, 𝑥 0] = 𝐾 (𝑥, 𝑥 0) + 𝑁 (𝑥, 𝑥 0),

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 25

 gaussian processes

covariance function for observation noise, 𝑁 where 𝑁 is the noise covariance: 𝑁 (𝑥, 𝑥 0) = cov[𝜀, 𝜀 0 | 𝑥, 𝑥 0]. The poste-
rior of the observation process is then a gp with

𝔼[𝑦 | 𝑥, D] = 𝜇 (𝑥) + 𝐶 (𝑥, x) (𝚺 + N) −1 (y − 𝝁);

(2.21)
cov[𝑦, 𝑦 0 | 𝑥, 𝑥 0, D] = 𝐶 (𝑥, 𝑥 0) − 𝐶 (𝑥, x) (𝚺 + N) −1𝐶 (x, 𝑥 0),

from which we can derive predictive distributions via (2.2).

2.3 overview of remainder of chapter

In the remainder of this chapter we will cover some additional, somewhat
niche and/or technical aspects of Gaussian processes that see occasional
20 In particular the claims regarding continuity use in Bayesian optimization. Modulo mathematical nuances irrelevant
and differentiability are slightly more compli- in practical settings, an intuitive (but not entirely accurate!) summary
cated than stated below.
follows:20
• a joint Gaussian process (discussed below) allows us to model multiple
multifidelity optimization: § 11.5, p. 263 related functions simultaneously, which is critical for some scenarios
multiobjective optimization: § 11.7, p. 269 such as multifidelity and multiobjective optimization;
sample path continuity: § 2.5, p. 28 • gp sample paths are continuous if the mean function is continuous and
the covariance function is continuous along the “diagonal” 𝑥 = 𝑥 0;
sample path differentiability: § 2.6, p. 30 • gp sample paths are differentiable if the mean function is differentiable
and the covariance function is differentiable along the “diagonal” 𝑥 = 𝑥 0;
derivative observations: § 2.6, p. 32 • a function with a sufficiently smooth gp distribution shares a joint gp
distribution with its gradient; among other things, this allows us to con-
dition on (potentially noisy) derivative observations via exact inference;
existence of global maxima: § 2.7, p. 34 • gp sample paths attain a maximum when sample paths are continuous
and the domain is compact,
uniqueness of global maxima: § 2.7, p. 34 • gp sample paths attain a unique maximum under the additional condition
that no two unique function values are perfectly correlated, and
inference with non-Gaussian observations • several methods are available for approximating the posterior process of
and constraints: § 2.8, p. 35 a gp conditioned on information incompatible with exact inference.
If satisfied with the above summary, the reader may safely skip this
material for now and move on with the next chapter. For those who wish
to see the gritty details, dive in below!

2.4 joint gaussian processes

In some settings, we may wish to reason jointly about two-or-more
related functions, such as an objective function and its gradient or an
expensive objective function and a cheaper surrogate. To this end we can
extend Gaussian processes to yield a joint distribution over the values
assumed by multiple functions. The key to the construction is to “paste
together” a collection of functions into a single function on a larger
domain, then construct a standard gp on this combined function.

26
 2.4. joint gaussian processes

Definition
To elaborate, consider a set of functions {𝑓𝑖 : X𝑖 → ℝ} we wish to
model.21 We define the disjoint union of these functions t𝑓 – defined on
Ã
disjoint union of {𝑓𝑖 }, t𝑓
the disjoint union22 of their domains X = X𝑖 – by insisting its restric- disjoint union of {X𝑖 }, X
tion to each domain be compatible with the corresponding function:
21 The domains need not be equal, but they often
t𝑓 : X → ℝ; t𝑓 | X𝑖 ≡ 𝑓𝑖 .
are in practice.

We now can define a gp on t𝑓 by choosing mean and covariance func- 22 A disjoint union represents a point 𝑥 ∈ X𝑖 by
the pair (𝑥, 𝑖), thereby combining the domains
tions on X as desired:
while retaining their identities.
𝑝 (t𝑓 ) = GP (t𝑓 ; 𝜇, 𝐾). (2.22)

We will call this construction a joint Gaussian process on {𝑓𝑖 }. joint Gaussian process
It is often convenient to decompose the moments of a joint gp into
their restrictions on relevant subspaces. For example, consider a joint gp
(2.22) on 𝑓 : F → ℝ and 𝑔 : G → ℝ. After defining

𝜇 𝑓 ≡ 𝜇| F ; 𝜇𝑔 ≡ 𝜇| G ;
𝐾 𝑓 ≡ 𝐾 | F ×F ; 𝐾𝑔 ≡ 𝐾 | G×G ; 𝐾 𝑓𝑔 ≡ 𝐾 | F ×G ; 𝐾𝑔𝑓 ≡ 𝐾 | G×F ,

we can see that 𝑓 and 𝑔 in fact have marginal gp distributions:23 23 In fact, any restriction of a gp-distributed func-
tion has a gp (or multivariate normal) distri-
𝑝 (𝑓 ) = GP (𝑓 ; 𝜇 𝑓 , 𝐾 𝑓 ); 𝑝 (𝑔) = GP (𝑔; 𝜇𝑔 , 𝐾𝑔 ), (2.23) bution.

that are coupled by the cross-covariance functions 𝐾 𝑓𝑔 and 𝐾𝑔𝑓 . Given

vectors x ⊂ F and x0 ⊂ G, these compute the covariance between the
corresponding function values 𝝓 = 𝑓 (x) and 𝜸 = 𝑔(x0):

𝐾 𝑓𝑔 (x, x0) = cov[𝝓, 𝜸 | x, x0];

(2.24)
𝐾𝑔𝑓 (x, x0) = cov[𝜸, 𝝓 | x, x0] = 𝐾 𝑓𝑔 (x, x0) >
.

When convenient we will notate a joint gp in terms of these decomposed 24 We also used this notation in (2.6), where the
functions, here writing:24 “domain” of the vector y can be taken to be
some finite index set of appropriate size.
     !
𝑓 𝜇𝑓 𝐾𝑓 𝐾 𝑓𝑔
𝑝 (𝑓, 𝑔) = GP ; , . (2.25)
𝑔 𝜇𝑔 𝐾𝑔𝑓 𝐾𝑔

With this notation, the marginal gp property (2.23) is perfectly analogous

to the marginal property of the multivariate Gaussian distribution (a.13).
We can also use this construction to define a gp on a vector-valued extension to vector-valued functions
function f: X → ℝ𝑑 by defining a joint Gaussian process on its 𝑑
coordinate functions {𝑓𝑖 } : X → ℝ. In this case we typically write the
resulting model using the standard notation GP (f; 𝜇, 𝐾), where the mean
and covariance functions are now understood to map to ℝ𝑑 and ℝ𝑑×𝑑 .

Example
We can demonstrate the behavior of a joint Gaussian process by extend-
ing our running example gp on 𝑓 : [0, 30] → ℝ. Recall the prior on 𝑓 has

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 27

gaussian processes

prior mean prior 95% credible interval sample

Figure 2.5: A joint Gaussian process over two functions on the shared domain X = [0, 30]. The marginal belief over both
functions is the same as our example gp from figure 2.1, but the cross-covariance (2.26) between the functions
strongly couples their behavior. We also show a sample from the joint distribution illustrating the strong
correlation induced by the joint prior.

zero mean function 𝜇 ≡ 0 and squared exponential covariance function

(2.4). We augment our original function with a companion function 𝑔,
defined on the same domain, that has exactly the same marginal gp
distribution. However, we couple the distribution of 𝑓 and 𝑔 by defining
a nontrivial cross-covariance function 𝐾 𝑓𝑔 (2.24):

𝐾 𝑓𝑔 (𝑥, 𝑥 0) = 0.9𝐾 (𝑥, 𝑥 0), (2.26)

where 𝐾 is the marginal covariance function of 𝑓 and 𝑔. A consequence

of this choice is that for any given point 𝑥 ∈ X, the correlation of the
corresponding function values 𝜙 = 𝑓 (𝑥) and 𝛾 = 𝑔(𝑥) is quite strong:

corr[𝜙, 𝛾 | 𝑥] = 0.9. (2.27)

We illustrate the resulting joint gp in figure 2.5. The marginal credible

intervals for 𝑓 (and now 𝑔) have not changed from our original example
in figure 2.1. However, drawing a sample of the functions from their joint
distribution reveals the strong coupling encoded in the prior (2.26–2.27).

Inference for joint Gaussian processes

The construction in (2.22) allows us to reason about a joint Gaussian
process as if it were a single gp. This allows us to condition a joint gp
on observations of jointly Gaussian distributed values following the
inference from jointly Gaussian distributed procedure outlined previously. In figure 2.6, we condition the joint gp
observations: § 2.2, p. 18 prior from figure 2.5 on ten observations: five exact observations of 𝑓 on
the left-hand side of the domain and five exact observations of 𝑔 on the
right-hand side. Due to the strong correlation between the two functions,
an observation of either function strongly informs our belief about the
other, even in regions where there are no direct observations.

2.5 continuity
In this and the following sections we will establish some important
properties of Gaussian processes determined by the properties of their

28
 2.5. continuity

observations (direct) posterior mean

observations (other function) posterior 95% credible interval

𝑝 (𝑓 | D)

𝑝 (𝑔 | D)

Figure 2.6: The joint posterior for our example joint gp prior in figure 2.5 conditioned on five exact observations of each
function.

moments. As a gp is completely specified by its mean and covariance

functions, it should not be surprising that the nature of these functions
have far-reaching implications regarding properties of the function being
modeled. A good familiarity with these implications can help guide model
design in practice – the focus of the next two chapters.
To begin, a fundamental question regarding Gaussian processes is
whether sample paths are almost surely continuous, and if so how many
times differentiable they may be. This is obviously an important consider-
ation for modeling and is also critical to ensure that global optimization existence of global maxima: § 2.7, p. 34
is a well-posed problem, as we will discuss later in this chapter. Fortu-
nately, continuity of Gaussian processes is a well-understood property
that can be guaranteed almost surely under simple conditions on the
mean and covariance functions.
Suppose 𝑓 : X → ℝ has distribution GP (𝑓 ; 𝜇, 𝐾). Recall that 𝑓 is
continuous at 𝑥 if 𝑓 (𝑥) − 𝑓 (𝑥 0) = 𝜙 − 𝜙 0 → 0 when 𝑥 0 → 𝑥. Continuity
is thus a limiting property of differences in function values. But under
the Gaussian process assumption, this difference is Gaussian distributed
(2.5, a.9)! We have
𝑝 (𝜙 − 𝜙 0 | 𝑥, 𝑥 0) = N (𝜙 − 𝜙 0; 𝑚, 𝑠 2 ),
where
𝑚 = 𝜇 (𝑥) − 𝜇 (𝑥 0); 𝑠 2 = 𝐾 (𝑥, 𝑥) − 2𝐾 (𝑥, 𝑥 0) + 𝐾 (𝑥 ,0 𝑥 0).
Now if 𝜇 is continuous at 𝑥 and 𝐾 is continuous at 𝑥 = 𝑥 ,0 then both
𝑚 → 0 and 𝑠 2 → 0 as 𝑥 → 𝑥 ,0 and thus 𝜙 −𝜙 0 converges in probability to
0. This intuitive condition of continuous moments is known as continuity continuity in mean square
in mean square at 𝑥; if 𝜇 and 𝐾 are both continuous over the entire domain
(the latter along the “diagonal” 𝑥 = 𝑥 0), then we say the entire process is
continuous in mean square.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 29

 gaussian processes

It turns out that continuity in mean square is not quite sufficient

to guarantee that 𝑓 is simultaneously continuous at every 𝑥 ∈ X with
sample path continuity probability one, a property known as sample path continuity. However,
very slightly stronger conditions on the moments of a gp are sufficient
25 r. j. adler and j. e. taylor (2007). Random to guarantee sample path continuity.25 The following result is adequate
Fields and Geometry. Springer–Verlag. [§§ 1.3– for most settings arising in practice and may be proven as a corollary to
1.4]
the slightly weaker (and slightly more complicated) conditions assumed
in adler and taylor’s theorem 1.4.1.
Theorem. Suppose X ⊂ ℝ𝑑 is compact and 𝑓 : X → ℝ has Gaussian
26 Hölder continuity is a generalization of Lip- process distribution GP (𝑓 ; 𝜇, 𝐾), where 𝜇 is continuous and 𝐾 is Hölder
schitz continuity. Effectively, the covariance continuous.26 Then 𝑓 is almost surely continuous on X.
function must, in some sense, be “predictably”
continuous. The condition that X ⊂ ℝ𝑑 be compact is equivalent to the domain
27 w. rudin (1976). Principles of Mathematical being closed and bounded, by the Heine–Borel theorem.27 Applying this
Analysis. McGraw–Hill. [theorem 2.41] result to our example gp in figure 2.1, we conclude that samples from the
process are continuous with probability one as the domain X = [0, 30] is
28 Following the discussion in the next section, compact and the squared exponential covariance function (2.4) is Hölder
they in fact are infinitely differentiable. continuous. Indeed, the generated samples are very smooth.28
Sample path continuity can also be guaranteed on non-Euclidean
domains under similar smoothness conditions.25

2.6 differentiability
We can approach the question of differentiability by again reasoning
about the limiting behavior of linear transformations of function values.
Suppose 𝑓 : X → ℝ with X ⊂ ℝ𝑑 has distribution GP (𝑓 ; 𝜇, 𝐾), and
consider the 𝑖th partial derivative of 𝑓 at x, if it exists:
𝜕𝑓 𝑓 (x + ℎe𝑖 ) − 𝑓 (x)
(x) = lim ,
𝜕𝑥𝑖 ℎ→0 ℎ
where e𝑖 is the 𝑖th standard basis vector. For ℎ > 0, the value in the limit
is Gaussian distributed as a linear transformation of Gaussian-distributed
random variables (a.9). Assuming the corresponding partial derivative
of the mean exists at x and the corresponding partial derivative with
respect to each input of the covariance function exists at x = x,0 then as
sequences of normal rvs: § a.2, p. 300 ℎ → 0 the partial derivative converges in distribution to a Gaussian:
   
𝜕𝑓 𝜕𝑓 𝜕𝜇 𝜕2𝐾
𝑝 (x) | x = N (x); (x), (x, x) .
𝜕𝑥𝑖 𝜕𝑥𝑖 𝜕𝑥𝑖 𝜕𝑥𝑖 𝜕𝑥𝑖0

differentiability in mean square If this property holds for each coordinate 1 ≤ 𝑖 ≤ 𝑑, then 𝑓 is said to be
differentiable in mean square at x.
joint gp between function and gradient If 𝑓 is differentiable in mean square everywhere in the domain, the
process itself is called differentiable in mean square, and we have the
remarkable result that the function and its gradient have a joint Gaussian
process distribution:
     !
𝑓 𝜇 𝐾 𝐾 ∇>
𝑝 (𝑓, ∇𝑓 ) = GP ; , . (2.28)
∇𝑓 ∇𝜇 ∇𝐾 ∇𝐾∇>
30
 2.6. differentiability

observations posterior mean posterior 95% credible interval

𝑑𝑓
𝑑𝑥

Figure 2.7: The joint posterior of the function and its derivative for our example Gaussian process from figure 2.2. The
dashed line in the lower plot corresponds to a derivative of zero.

Here by writing the gradient operator ∇ on the left-hand side of 𝐾 we

mean the result of taking the gradient with respect to its first input,
and by writing ∇> on the right-hand side of 𝐾 we mean taking the
gradient with respect to its second input and transposing the result. Thus
∇𝐾 : X × X → ℝ𝑑 maps pairs of points to column vectors: covariance between ∇𝑓 (x) and 𝑓 (x0 ), ∇𝐾
 
  𝜕𝑓 
0  𝜕𝐾
∇𝐾 (x, x ) 𝑖 = cov
0 0
(x), 𝑓 (x ) x, x = (x, x0),
𝜕𝑥𝑖 𝜕𝑥𝑖

and 𝐾 ∇>: X × X → (ℝ𝑑 ) ∗ maps pairs of points to row vectors: transpose of covariance between 𝑓 (x) and
 
∇𝑓 (x0 ), 𝐾 ∇>
𝐾∇>(x, x0) = ∇𝐾 (x,0 x) > .

Finally, the function ∇𝐾∇>: X × X → ℝ𝑑×𝑑 represents the result of covariance between ∇𝑓 (x) and ∇𝑓 (x0 ),
applying both operations, mapping a pair of points to the covariance ∇𝐾 ∇>
matrix between the entries of the corresponding gradients:
 
  𝜕𝑓 𝜕𝑓 0  𝜕2𝐾
∇𝐾∇ (x, x ) 𝑖 𝑗 = cov
> 0 0
(x), 0 (x ) x, x = (x, x0).
𝜕𝑥𝑖 𝜕𝑥𝑗 𝜕𝑥𝑖 𝜕𝑥𝑗0

As the gradient of 𝑓 has a Gaussian process marginal distribution (2.28),

we can reduce the question of continuous differentiability to sample path continuous differentiability
continuity of the gradient process following the discussion above.
Figure 2.7 shows the posterior distribution for the derivative of our example and discussion
example Gaussian process alongside the posterior for the function itself.
We can observe a clear correspondence between the two distributions;
for example, the posterior mean of the derivative vanishes at critical
points of the posterior mean of the function. Notably, we have a great
deal of residual uncertainty about the derivative, even at the observed
locations. That is because the relatively high spacing between the exist-
ing observations limits our ability to accurately estimate the derivative

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 31

 gaussian processes

observations posterior mean posterior 95% credible interval

𝑑𝑓
𝑑𝑥

Figure 2.8: The joint posterior of the derivative of our example Gaussian process after adding a new observation nearby
another suggesting a large positive slope. The dashed line in the lower plot corresponds to a derivative of zero.

anywhere. Adding an observation immediately next to a previous one

significantly reduces the uncertainty in the derivative in that region by
effectively providing a finite-difference approximation; see figure 2.8.

Conditioning on derivative observations

However, we can be more direct in specifying derivatives than finite
differencing. We can instead condition the joint gp (2.28) directly on a
inference from jointly Gaussian distributed derivative observation, as described previously. Figure 2.9 shows the joint
observations: § 2.2, p. 18 posterior after conditioning on an exact observation of the derivative
at to the left-most observation location, where the uncertainty in the
derivative now vanishes entirely. This capability allows the seamless
incorporation of derivative information into an objective function model.
Notably, we can even condition a Gaussian process on noisy derivative
observations as well, as we might obtain in stochastic gradient descent.
29 For 𝐾 we again only need to consider the “di- We can reason about derivatives past the first recursively. For exam-
agonal” x = x0. ple, if 𝜇 and 𝐾 are twice differentiable,29 then the (e.g., half-vectorized30)
30 Recall the Hessian is symmetric (assuming the Hessian of 𝑓 will also have a joint gp distribution with 𝑓 and its gradient.
second partial derivatives are continuous) and Defining h to be the operator mapping a function to its half-vectorized
thus redundant. The half-vectorization opera-
tor vech A maps the upper triangular part of
Hessian:
a square, symmetric matrix A to a vector. h𝑓 = vech ∇∇>𝑓,
for a Gaussian process with suitably differentiable moments, we have


𝑝 (h𝑓 ) = GP h𝑓 ; h𝜇, h𝐾 h> , (2.29)
where we have used the same notational convention for the transpose.
Further, 𝑓, ∇𝑓, and h𝑓 will have a joint Gaussian process distribution
given by augmenting (2.28) with the marginal in (2.29) and the cross-
covariance functions
cov[h𝑓, 𝑓 ] = h𝐾; cov[h𝑓, ∇𝑓 ] = h𝐾 ∇.>

32
 2.7. existence and uniqeness of global maxima

observations posterior mean posterior 95% credible interval

𝑑𝑓
𝑑𝑥

Figure 2.9: The joint posterior of the derivative of our example Gaussian process after adding an exact observation of the
derivative at the indicated location. The dashed line in the lower plot corresponds to a derivative of zero.

We can continue further in this vein if needed; however, we rarely reason

about derivatives of third-or-higher order in Bayesian optimization.31 31 This is true in classical optimization as well!

Other linear transformations

The joint gp distribution between a suitably smooth gp-distributed func-
tion and its gradient (2.28) is simply an infinite-dimensional analog of
the general result that Gaussian random variables are jointly Gaussian
distributed with arbitrary linear transformations (a.10), after noting that
differentiation is a linear operator. We can extend this result to reason
about other linear transformations of gp-distributed functions. diaco-
nis’s original motivation for studying Bayesian numerical methods was
quadrature, the numerical estimation of intractable integrals.32 It turns 32 p. diaconis (1988). Bayesian Numerical Anal-
out that Gaussian processes are a rather convenient model for this task: ysis. In: Statistical Decision Theory and Related
Topics iv.
if 𝑝 (𝑓 ) = GP (𝑓 ; 𝜇, 𝐾) and we want to reason about the expectation
∫
𝑍 = 𝑓 (𝑥) 𝑝 (𝑥) d𝑥,

then (under mild conditions) we again have a joint Gaussian process 33 This can be shown, for example, by consider-
distribution over 𝑓 and 𝑍 .33 This enables both inference about 𝑍 and ing the limiting distribution of Riemann sums.
conditioning on noisy observations of integrals, such as a Monte Carlo 34 a. o’hagan (1991). Bayes–Hermite quadrature.
estimate of an expectation. The former is the basis for Bayesian quadra- Journal of Statistical Planning and Inference
29(3):245–260
ture, an analog of Bayesian optimization bringing Bayesian experimental
design to bear on numerical integration.32, 34, 35 35 c. e. rasmussen and z. ghahramani (2002).
Bayesian Monte Carlo. neurips 2002

2.7 existence and uniqeness of global maxima

The primary use of gps in Bayesian optimization is to inform optimiza-
tion decisions, which will be our focus for the majority of this book.
Before continuing down this path, we pause to consider whether global

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 33

 gaussian processes

optimization of a gp-distributed function is a well-posed problem, in par-

ticular, whether the model guarantees the existence of a global maximum
at all.
Consider a function 𝑓 : X → ℝ with distribution GP (𝑓 ; 𝜇, 𝐾), and
consider the location and value of its global optimum, if one exists:

𝑥 ∗ = arg max 𝑓 (𝑥); 𝑓 ∗ = max 𝑓 (𝑥) = 𝑓 (𝑥 ∗ ).

𝑥 ∈X 𝑥 ∈X

mutual information and entropy search: § 7.6, As 𝑓 is unknown, these quantities are random variables. Many Bayesian
p. 135 optimization algorithms operate by reasoning about the distributions of
(and uncertainties in) these quantities induced by our belief on 𝑓.
There are two technical issues we must address. The first is whether
we can be certain that a globally optimal value 𝑓 ∗ exists when the objec-
tive function is random. If existence is not guaranteed, then its distribu-
tion is meaningless. The second issue is one of uniqueness: assuming the
objective does attain a maximal value, can we be certain the optimum
is unique? In general 𝑥 ∗ is a set-valued random variable, and thus its
distribution might have support over arbitrary subsets of the domain,
rendering it complicated to reason about. However, if we could ensure
the uniqueness of 𝑥 ∗, its distribution would have support on X rather
than its power set, allowing more straightforward inference.
Both the existence of 𝑓 ∗ and uniqueness of 𝑥 ∗ are tacitly assumed
throughout the Bayesian optimization literature when building algo-
rithms based on distributions of these quantities, but these properties
are not guaranteed for arbitrary Gaussian processes. However, we can
ensure these properties hold almost surely under mild conditions.

Existence of global maxima

To begin, guaranteeing the existence of an optimal value is straightfor-
ward if we suppose the domain X is compact, a pervasive assumption
in optimization. This is no coincidence! In this case, if 𝑓 is continuous
36 w. rudin (1976). Principles of Mathematical then it achieves a global optimum by the extreme value theorem.36 Thus
Analysis. McGraw–Hill. [theorem 4.16] sample path continuity of 𝑓 and compactness of X is sufficient to ensure
that 𝑓 ∗ exists almost surely. Both conditions can be readily established:
sample path continuity: § 2.5, p. 28 sample path continuity by following our previous discussion, and com-
pactness of the domain by standard arguments (for example, ensuring
that X ⊂ ℝ𝑑 be closed and bounded).

Uniqueness of global maxima

We now turn to the question of uniqueness of 𝑥 ,∗ which obviously only
becomes a meaningful question after presupposing that 𝑓 ∗ exists. Again,
37 A centered Gaussian process has identaically this condition is easy to ensure almost surely under simple conditions
zero mean function 𝜇 ≡ 0. on the covariance function of a Gaussian process.
38 j. kim and d. pollard (1990). Cube Root kim and pollard considered this issue and provided straightforward
Asymptotics. The Annals of Statistics 18(1):191– conditions under which the uniqueness of 𝑥 ∗ is guaranteed for a centered
219. [lemma 2.6] Gaussian process.37, 38 Namely, no two unique points in the domain can

34
 2.8. inference with non-gaussian observations and constraints

have perfectly correlated function values, a natural condition that can

be easily verified.
Theorem (kim and pollard, 1990). Let X be a compact metric space.39 39 Although unlikely to matter in practice, kim
Suppose 𝑓 : X → ℝ has distribution GP (𝑓 ; 𝜇 ≡ 0, 𝐾), and that 𝑓 is sample and pollard allow X to be 𝜎-compact and
show that the supremum (rather than the max-
path continuous. If for all 𝑥, 𝑥 0 ∈ X with 𝑥 ≠ 𝑥 0 we have imum) is unique under the same conditions.

var[𝜙 − 𝜙 0 | 𝑥, 𝑥 0] = 𝐾 (𝑥, 𝑥) − 2𝐾 (𝑥, 𝑥 0) + 𝐾 (𝑥 ,0 𝑥 0) ≠ 0,

then 𝑓 almost surely has a unique maximum on X.

arcones provided slightly weaker conditions for uniqueness of the 40 m. a. arcones (1992). On the arg max of a Gaus-
supremum, avoiding the requirement of sample path continuity.40 sian process. Statistics & Probability Letters
15(5):373–374.

Counterexamples
Although the above conditions for ensuring existence of 𝑓 ∗ and unique-
ness of 𝑥 ∗ are fairly mild, it is easy to construct counterexamples.
Consider a function on the closed unit interval, which we note is 41 It turns out this naïve model of white noise
compact: 𝑓 : [0, 1] → ℝ. We endow 𝑓 with a “white noise”41 Gaussian has horrible mathematical properties, but it is
sufficient for this counterexample.
process with
𝜇 (𝑥) ≡ 0; 𝐾 (𝑥, 𝑥 0) = [𝑥 = 𝑥 0].
Now 𝑓 almost surely does not have a maximum. Roughly, because the 42 Let 𝑄 = ℚ ∩ [0, 1] = {𝑞𝑖 } be the rationals in
value of 𝑓 at every point in the domain is independent of every other, the domain and let 𝑓 ∗ be a putative maximum.
Defining 𝜙𝑖 = 𝑓 (𝑞𝑖 ), we must have 𝜙𝑖 ≤ 𝑓 ∗
there will almost always be a point with value exceeding any putative for every 𝑖; call this event 𝐴.
maximum.42 However, the conditions of sample path continuity were Define the event 𝐴𝑘 by 𝑓 ∗ exceeding the
violated as the covariance is discontinuous at 𝑥 = 𝑥 .0 first 𝑘 elements of 𝑄. From independence,
We may also construct a Gaussian process that almost surely achieves Ö
𝑘
a maximum that is not unique. Consider a random function 𝑓 defined Pr(𝐴𝑘 ) = Pr(𝜙𝑖 ≤ 𝑓 ∗ ) = Φ(𝑓 ∗ ) 𝑘,
on the (compact) interval [0, 4𝜋] defined by the parametric model 𝑖=1

so Pr(𝐴𝑘 ) → 0 as 𝑘 → ∞. But {𝐴𝑘 } % 𝐴,

𝑓 (𝑥) = 𝛼 cos 𝑥 + 𝛽 sin 𝑥, so Pr(𝐴) = 0, and 𝑓 ∗ is almost surely not the
maximum.
where 𝛼 and 𝛽 are independent standard normal random variables. Then
𝑓 has a Gaussian process distribution with

𝜇 (𝑥) ≡ 0; 𝐾 (𝑥, 𝑥 0) = cos(𝑥 − 𝑥 0). (2.30)

Here 𝜇 is continuous and 𝐾 is Hölder continuous, and thus 𝑓 is sample

path continuous and almost surely achieves a global maximum. However,
𝑓 is also periodic with period 2𝜋 with probability one and will thus almost
surely achieve its maximum twice. Note that the covariance function does Our counterexample gp without a unique max-
not satisfy the conditions outlined in the above theorem, as any input imum. Every sample achieves its maximum
twice.
locations separated by 2𝜋 have perfectly correlated function values.

2.8 inference with non-gaussian observations and constraints

Gaussian process inference is tractable when the observed values are inference from jointly Gaussian distributed
jointly Gaussian distributed with the function of interest (2.6). However, observations: § 2.2, p. 18
this may not always hold for all relevant information we may receive.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 35

 gaussian processes

observations objective mean, Gaussian noise 95% credible interval, Gaussian noise

Figure 2.10: Regression with observations corrupted with heavy-tailed noise. The triangular marks indicate observations
lying beyond the plotted range. Shown is the posterior distribution of an objective function (ground truth
plotted in red) modeling the errors as Gaussian. The posterior is heavily affected by the outliers.

𝑝 (𝑦 | 𝑥, 𝜙)
One obvious limitation is an incompatibility with naturally non-
Gaussian observations. A scenario particularly relevant to optimization
is heavy-tailed noise. Consider the data shown in figure 2.10, where
some observations represent extreme outliers. These errors are poorly
modeled as Gaussian, and attempting to infer the underlying objective
𝑦 function with the additive Gaussian noise model leads to overfitting
and poor predictive performance. A Student-𝑡 error model with 𝜈 ≈ 4
𝜙 degrees of freedom provides a robust alternative:43
A Student-𝑡 error model (solid) with a Gaus-
sian error model (dashed) for reference. The 𝑝 (𝑦 | 𝑥, 𝜙) = T (𝑦; 𝜙, 𝜎𝑛2 , 𝜈). (2.31)
heavier tails of the Student-𝑡 model can better
explain large outliers. The heavier tails of this model can better explain large outliers; un-
fortunately, the non-Gaussian nature of this model also renders exact
43 k. l. lange et al. (1989). Robust Statistical Mod- inference impossible. We will demonstrate how to overcome this impasse.
eling Using the 𝑡 Distribution. Journal of the Constraints on an objective function, such as bounds on given func-
American Statistical Association 84(408):881–
896. tion values, can also provide valuable information during optimization,
but many natural constraints cannot be reduced to observations that can
be handled in closed form. Several Bayesian optimization policies impose
hypothetical constraints on the objective function when designing each
observation, requiring inference from intractable constraints even when
the observations themselves pose no difficulties.
To see how constraints might arise in optimization, consider a Gaus-
sian process belief on a one-dimensional objective 𝑓, and suppose we
wish to condition on 𝑓 on having a local maximum at a given loca-
differentiability, derivative observations: § 2.6, tion 𝑥. Assuming the function is twice differentiable, we can invoke the
p. 30 second-derivative test to encode this information in two constraints:

𝑓 0 (𝑥) = 0; 𝑓 00 (𝑥) < 0. (2.32)

We can condition a gp on the first of these conditions by following

our previous discussion. However, no gp is compatible with the second

36
 2.8. inference with non-gaussian observations and constraints

true distribution
samples Figure 2.11: The probability density function of an example
distribution along with 50 samples drawn inde-
pendently from the distribution. In Monte Carlo
approaches, the distribution is effectively approxi-
mated by a mixture of Dirac delta distributions at
the sample locations.

condition as 𝑓 00 (𝑥) would necessarily have a Gaussian distribution with

unbounded support (2.29). We need some other means to proceed.

Non-Gaussian observations: general case

We can address both non-Gaussian observations and constraints with
the following general case, which is flexible enough to handle a large
range of information. As in our discussion on exact inference, suppose
there is some vector y sharing a joint Gaussian process distribution with
a function of interest 𝑓 (2.6):
     !
𝑓 𝜇 𝐾 𝜅>
𝑝 (𝑓, y) = GP ; , .
y m 𝜅 C

Suppose we receive some information about y in the form of infor-

mation D inducing a non-Gaussian posterior on y. Here, it is convenient
to adopt the language of factor graphs44 and write the resulting posterior 44 f. r. kschischang et al. (2001). Factor Graphs
as proportional to the prior weighted by a function 𝑡 (y) encoding the and the Sum–Product Algorithm. ieee Trans-
actions on Information Theory 47(2):498–519.
available information, which may factorize:
Ö
𝑝 (y | D) ∝ 𝑝 (y) 𝑡 (y) = N (y; m, C) 𝑡𝑖 (y). (2.33)
𝑖

The functions {𝑡𝑖 } are called factors or local functions that may comprise factors, local functions, {𝑡𝑖 }
a likelihood augmented by any desired (hard or soft) constraints. The
term “local functions” arises because each factor often depends only on 45 For example, when observations are condi-
a low-dimensional subspace of y, often a single entry.45 tionally independent given the corresponding
function values, the likelihood factorizes into
The posterior on y (2.33) in turn induces a posterior on 𝑓 : a product of one-dimensional factors (1.3):
∫ Ö
𝑝 (𝑓 | D) = 𝑝 (𝑓 | y) 𝑝 (y | D) dy. (2.34) 𝑝 (y | x, 𝝓) = 𝑝 (𝑦𝑖 | 𝑥𝑖 , 𝜙𝑖 ).
𝑖

At first glance, we may hope to resolve this posterior easily as 𝑝 (𝑓 | y) is

a Gaussian process (2.9–2.10). Unfortunately, the non-Gaussian posterior
on y usually renders the posterior on 𝑓 intractable.

Monte Carlo sampling

A Monte Carlo approach to approximating the 𝑓 posterior (2.34) begins
by drawing samples from the y posterior (2.33):
{y𝑖 }𝑠𝑖=1 ∼ 𝑝 (y | D).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 37

 gaussian processes

observations objective posterior mean posterior 95% credible interval

Figure 2.12: Regression with observations corrupted with heavy-tailed noise. The triangular marks indicate observations
lying beyond the plotted range. Shown is the posterior distribution of an objective function (ground truth
plotted in red) modeling the errors as Student-𝑡 distributed with 𝜈 = 4 degrees of freedom. The posterior was
approximated from 100 000 Monte Carlo samples. Comparing with the additive Gaussian noise model from
figure 2.10, this model effectively ignores the outliers and the fit is excellent.

We may generate these by appealing to one of numerous Markov chain

46 Handbook of Markov Chain Monte Carlo (2011). Monte Carlo (mcmc) routines.46 One natural choice would be elliptical
Chapman & Hall. slice sampling,47 which is specifically tailored for latent Gaussian models
47 i. murray et al. (2010). Elliptical slice sampling. of this form. Samples from a one-dimensional toy example distribution
aistats 2010. are shown in figure 2.11.
Given posterior samples of y, we may then approximate (2.34) via
the standard Monte Carlo estimator

1 ∑︁ 1 ∑︁
𝑠 𝑠
𝑝 (𝑓 | D) ≈ 𝑝 (𝑓 | y𝑖 ) = GP (𝑓 ; 𝜇D𝑖 , 𝐾D ). (2.35)
𝑠 𝑖=1 𝑠 𝑖=1

This is a mixture of Gaussian processes, each of the form in (2.9–2.10).

The posterior mean functions depend on the corresponding y samples,
whereas the posterior covariance functions are identical as there is no
dependence on the observed values. In this approximation, the marginal
belief about any function value is then a mixture of univariate Gaussians:

1 ∑︁
𝑠
𝑝 (𝜙 | 𝑥, D) ≈ N (𝜙; 𝜇𝑖 , 𝜎 2 ); 𝜇𝑖 = 𝜇D𝑖 (𝑥); 𝜎 2 = 𝐾D (𝑥, 𝑥).
𝑠 𝑖=1
(2.36)
Although slightly more complex than the Gaussian marginals of a Gaus-
sian process, this is often convenient enough for most needs.
example: Student-𝑡 observation model A Monte Carlo approximation to the posterior for the heavy-tailed
dataset from figure 2.10 is shown in figure 2.12. The observations were
modeled as corrupted by Student-𝑡 errors with 𝜈 = 4 degrees of freedom.
The posterior was approximated using a truly excessive number of sam-
ples (100 000, with a burn-in of 10 000) from the y posterior drawn using
elliptical slice sampling.47 The outliers in the data are ignored and the
predictive performance is excellent.

38
 2.8. inference with non-gaussian observations and constraints

Gaussian approximate inference

An alternative to sampling is approximate inference, where we make
a parametric approximation to the y posterior that yields a tractable
posterior on 𝑓. In particular, if the posterior (2.33) were actually normal,
it would induce a Gaussian process posterior on 𝑓. This insight is the
basis for most approximation schemes.
In this vein, we proceed by first – somehow – approximating the
true posterior over y with a multivariate Gaussian distribution:

𝑝 (y | D) ≈ 𝑞(y | D) = N (y; m̃, C̃). (2.37)

We are free to design this approximation as we see fit. There are several
general-purpose approaches available, distinguished by how they ap-
proach maximizing the fidelity of fitting the true posterior (2.33). These
include the Laplace approximation, Gaussian expectation propagation, Laplace approximation: § b.1, p. 301
and variational Bayesian inference. The first two of these methods are Gaussian expectation propagation: § b.2
covered in appendix b, and nickisch and rasmussen provide an exten- p. 302
sive survey of these and other approaches in the context of Gaussian
process binary classification.48 48 h. nickisch and c. e. rasmussen (2008). App-
Regardless of the details of the approximation scheme, the high-level proximations for Binary Gaussian Process
Classification. Journal of Machine Learning Re-
result is the same – the normal approximation (2.37) in turn induces an search 9(Oct):2035–2078.
approximate Gaussian process posterior on 𝑓. To demonstrate this, we
consider the posterior on 𝑓 that would arise from a direct observation of
y (2.9–2.10) and integrate against the approximate posterior (2.37):
∫
𝑝 (𝑓 | D) ≈ 𝑝 (𝑓 | y) 𝑞(y | D) dy = GP (𝑓 ; 𝜇D , 𝐾D ), (2.38)
where
𝜇D (𝑥) = 𝜇 (𝑥) + 𝜅 (𝑥)> C−1 ( m̃ − m);
(2.39)
𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − 𝜅 (𝑥)> C−1 (C − C̃) C−1 𝜅 (𝑥 0).

For most approximation schemes, the posterior covariance on 𝑓

simplifies to a nicer, more familiar form. Most approximations to the y
posterior (2.37) yield an approximate posterior covariance of the form

C̃ = C − C (C + N) −1 C, (2.40)

where N is positive definite. Although this might appear mysterious, it

is actually a natural form: it is the posterior covariance that would result
from observing y corrupted by additive Gaussian noise with covariance N
(2.19), except we are now free to design the noise covariance to maximize
the fit. For approximations of this form (2.40), the approximate posterior
covariance function on 𝑓 simplifies to the more familiar

𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − 𝜅 (𝑥)> (C + N) −1 𝜅 (𝑥 0). (2.41)

To demonstrate the power of approximate inference, we return to example: conditioning on a local optimum
our motivating scenario of conditioning a one-dimensional process on
having a local maximum at an identified point 𝑥, which we can achieve by

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 39

 gaussian processes

posterior mean posterior 95% credible interval posterior sample

𝑥 𝑥 𝑥

𝑓 0 (𝑥) = 0 𝑓 00 (𝑥) < 0

Figure 2.13: Approximately conditioning a Gaussian process to have a local maximum at the marked point 𝑥. We show
each stage of the conditioning process with a sample drawn from the corresponding posterior. We begin
with the unconstrained process (left), which we condition on the first derivative being zero at 𝑥 using exact
inference (middle). Finally we use Gaussian expectation propagation to approximately condition on the second
derivative being negative at 𝑥.

conditioning the first derivative to be zero and constraining the second

derivative to be negative at 𝑥 (2.32). We illustrate an approximation to
the resulting posterior step-by-step in figure 2.13, beginning with the
example Gaussian process in the left-most panel. We first condition
derivative observations: § 2.6, p. 32 the process on the first derivative observation 𝑓 0 (𝑥) = 0 using exact
inference; the result is shown in the middle panel. Both the updated
posterior mean and the sample reflect this information; however, the
sample displays a local minimum at 𝑥, as the second-derivative constraint
has not yet been addressed.
To incorporate the second-derivative constraint, we begin with this
updated gp and consider the second derivative ℎ = 𝑓 00 (𝑥), which is
Gaussian distributed prior to the constraint (2.29):

𝑝 (ℎ) = N (ℎ; 𝑚, 𝑠 2 ).

The negativity constraint induces a posterior on ℎ incorporating the

factor [ℎ < 0] (2.33); see figure 2.14:

𝑝 (ℎ | D) ∝ 𝑝 (ℎ) [ℎ < 0].

The result is a truncated normal posterior on ℎ. We may use Gaussian

expectation propagation, which is especially convenient for handling
bound constraints of this form, to produce a Gaussian approximation:

˜ 𝑠˜2 ).
𝑝 (ℎ | D) ≈ 𝑞(ℎ | D) = N (ℎ; 𝑚,

Incorporating the updated belief on ℎ into the Gaussian process (2.39)

yields the approximate posterior in the right-most panel of figure 2.13.
Although there is still some residual probability that the second derivative
49 p. mccullagh and j. a. nelder (1989). Gener- is positive at 𝑥 in the approximate posterior (approximately 8%; see figure
alized Linear Models. Chapman & Hall. 2.14), the belief reflects the desired information reasonably faithfully.

40
 2.9. summary of major ideas

prior, 𝑝 (ℎ) true posterior, 𝑝 (ℎ | D) ∝ 𝑝 (ℎ) [ℎ < 0]

constraint factor, [ℎ < 0] Gaussian ep approximation

0 0

Figure 2.14: A demonstration of Gaussian expectation propagation. On the left we have a Gaussian belief on the second
derivative, 𝑝 (ℎ). We wish to constrain this value to be negative, introducing a step-function factor encoding
the constraint, [ℎ < 0]. The resulting distribution is non-Gaussian (right), but we can approximate it with a
Gaussian, which induces an updated gp posterior on the function approximately incorporating the constraint.

Going beyond this example, we may use the approach outlined above 50 h. nickisch and c. e. rasmussen (2008). App-
to realize a general framework for Bayesian nonlinear regression by proximations for Binary Gaussian Process
Classification. Journal of Machine Learning Re-
combining a gp prior on a latent function with an observation model search 9(Oct):2035–2078.
appropriate for the task at hand, then approximating the posterior as
51 j. møller et al. (1998). Log Gaussian Cox Pro-
desired. The convenience and modeling flexibility offered by Gaussian cesses. Scandinavian Journal of Statistics 25(3):
processes can easily justify any extra effort required for approximating 451–482.
the posterior. This can be seen as a nonlinear extension of the well-known 52 r. p. adams et al. (2009). Tractable Nonpara-
family of generalized linear models.49 metric Bayesian Inference in Poisson Pro-
This approach is quite popular and been realized countless times. cesses with Gaussian Process Intensities. icml
Notable examples include binary classification using a logistic or probit 2009.

observation model,50 modeling point processes as a nonhomogenous 53 m. kuss (2006). Gaussian Process Models for
Robust Regression, Classification, and Rein-
Poisson process with unknown intensity,51, 52 and robust regression with forcement Learning. PhD thesis. Technische
heavy-tailed additive noise such as Laplace53 or Student-𝑡54, 55 distributed Universität Darmstadt.[§ 5.4]
errors. With regard to the latter and our previous heavy-tailed noise 54 r. m. neal (1997). Monte Carlo Implementation
example, a Laplace approximation to the posterior for the data in figures of Gaussian Process Models for Bayesian Re-
2.10–2.12 with the Student-𝑡 observation model produces an approximate gression and Classification. Technical report
posterior in excellent agreement with the Monte Carlo approximation (9702). Department of Statistics, University of
Toronto.
in figure 2.12; see figure 2.15. The cost of approximate inference in this
55 p. jylänki et al. (2011). Robust Gaussian Pro-
case was dramatically (several orders of magnitude) cheaper than Monte cess Regression with a Student-𝑡 Likelihood.
Carlo sampling. Journal of Machine Learning Research 12(99):
3227–3257.
56 diaconis identified an early application of gps
2.9 summary of major ideas by poincaré for nonlinear regression:
Gaussian processes have been studied – in one form or another – for over p. diaconis (1988). Bayesian Numerical Anal-
100 years.56 Although we have covered a lot of ground in this chapter, we ysis. In: Statistical Decision Theory and Related
have only scratched the surface of an expansive body of literature. A good Topics iv.

entry point to that literature is rasmussen and williams’s monograph, h. poincaré (1912). Calcul des probabilités.
Gauthier–Villars.
which focuses on machine learning applications of Gaussian processes
but also covers their theoretical underpinnings and properties in depth.57 57 c. e. rasmussen and c. k. i. williams (2006).
Gaussian Processes for Machine Learning. mit
A good companion to this work is the book of adler and taylor, which Press.
takes a deep dive into the properties and geometry of sample paths, 58 r. j. adler and j. e. taylor (2007). Random
including statistical properties of their maxima.58 Fields and Geometry. Springer–Verlag.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 41

gaussian processes

observations objective posterior mean posterior 95% credible interval

Figure 2.15: A Laplace approximation to the posterior from figure 2.12.

Fortunately, the basic definitions and properties covered in § 2.1 and

exact inference procedure covered in § 2.2 already provide a sufficient
foundation for the majority of practical applications of Bayesian opti-
mization. This material also provides sufficient background knowledge
for the majority of the remainder of the book. However, we wish to
underscore the major results from this chapter at a high level.
• Gaussian processes extend the multivariate normal distribution to model
functions on infinite domains. As in the finite-dimensional case, Gaussian
processes are specified by their first two moments – a mean function
and a positive-definite covariance function – which endow any finite set
of function values with a multivariate normal distribution (2.2–2.3).
• Conditioning a Gaussian process on function observations that are either
exact or corrupted by additive Gaussian noise yields a Gaussian process
posterior with updated moments reflecting the assumptions in the prior
and the information in the observations (2.9–2.10).
inference from arbitrary joint Gaussian • In fact, we may condition a Gaussian process on the observation of any
observations: § 2.2, p. 22 observations sharing a joint Gaussian distribution with the function of
interest.
interpretation of posterior moments: § 2.2, • In the case of exact inference, the posterior moments of a Gaussian
p. 21 process can be rewritten in terms of correlations among function values
and 𝑧-scores of the observed values in a manner that may be more
intuitive than the standard formulas.
joint Gaussian processes: § 2.4, p. 26 • We may extend Gaussian processes to jointly model multiple correlated
functions via careful bookkeeping, a construction known as a joint Gaus-
extensions and other settings: chapter 11, sian process. Joint gps are widely used in optimization settings involving
p. 245 multiple objectives and/or cheaper surrogates for an expensive objective.
continuity: § 2.5, p. 28 • Continuity and differentiability of Gaussian process sample paths can
differentiability: § 2.6, p. 30 be guaranteed under mild assumptions on the mean and covariance
functions. When these functions are sufficiently differentiable, a gp-
distributed function shares a joint gp distribution with its gradient (2.28).

42
 2.9. summary of major ideas

This joint distribution allows us to condition a Gaussian process on derivative observations: § 2.6, p. 32
(potentially noisy) derivative observations.
• The existence and uniqueness of global maxima for Gaussian process existence and uniqueness of global maxima:
sample paths can be guaranteed under mild assumptions on the mean § 2.7, p. 33
and covariance functions. Establishing these properties ensures that
the location 𝑥 ∗ and value 𝑓 ∗ of the global maximum are well-founded 59 In particular, policies grounded in information
random variables, which will be critical for some optimization methods theory under the umbrella of “entropy search.”
See § 7.6, p. 135 for more.
introduced later in the book.59
• Inference from non-Gaussian observations and constraints is possible inference with non-Gaussian observations
via Monte Carlo sampling or Gaussian approximate inference. and constraints: § 2.8, p. 35

Looking forward, the focus of this chapter has been on theoretical

rather than practical properties of Gaussian processes. A huge outstand-
ing question is how to actually design a Gaussian process to model a
given system. This will be our focus for the next two chapters. In the
next chapter, we will explore model construction, and in the following
chapter we will consider model assessment in light of available data.
Finally, we have not yet discussed any computational issues inherent
to Gaussian process inference, including, most importantly, how the
cost of computing the posterior grows with respect to the number of
observations. We will discuss implementation details and scaling in a implementation and scaling of Gaussian
dedicated chapter later in the book. process inference: chapter 9, p. 201

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 43

 3
MODELING WI TH GAUSSIAN PROCESSES

Bayesian optimization relies on a faithful model of the system of in-

terest to make well-informed decisions. In fact, even more so than the
details of the optimization policy, the fidelity of the underlying model of
the objective function is the most decisive factor determining optimiza-
tion performance. This has been long acknowledged, with mockus for
example commenting in his seminal work that:1

The development of some system of a priori distributions suitable 1 j. mockus (1974). On Bayesian Methods for
for different classes of the function 𝑓 is probably the most important Seeking the Extremum. Optimization Tech-
niques ifip Technical Conference.
problem in the application of [the] Bayesian approach to. . . global
optimization.

The importance of careful modeling has not waned in the intervening

years, but our capacity for building sophisticated models has improved.
Recall our approach to modeling observations obtained during opti- Bayesian inference of the objective function:
mization combines a prior process for a (perhaps not directly observable) § 1.2, p. 8
objective function (1.8) and an observation model linking the values of
the objective to measured values (1.2). Both distributions must be speci-
fied before we can derive a posterior belief about the objective function
(1.10) and predictive distribution for proposed observations (1.7), which
together serve as the key enablers of Bayesian optimization policies.
In practice, the choice of observation model is often noncontrover-
sial,2 and our running prototypes of exact observation and additive 2 However, we may not be certain about some
Gaussian noise suffice for many systems. The bulk of modeling effort details, such as the scale of observation noise,
an issue we will address in the next chapter.
is thus spent crafting the prior process. Although specifying a Gaus-
sian process is seemingly as simple as choosing a mean and covariance
function, it can be difficult to intuit appropriate choices without a great
deal of knowledge about the system of interest. As an alternative to
prior knowledge, we may appeal to a data-driven approach, where we
establish a space of candidate models and search through this space for
those offering the best explanation of available data. Almost all Gaussian
process models used in practice are designed in this manner, and we will
lay the groundwork for this approach in this chapter and the next.
As a Gaussian process is specified by its first two moments, data-
driven model design boils down to searching for the prior mean and
covariance functions most harmonious with our observations. This can
be a daunting task as the space of possibilities is limitless. However, we do
not need to begin from scratch: there are mean and covariance functions
available off-the-shelf for modeling a variety behavioral archetypes, and
by systematically combining these components we may model functions
with a rich variety of behavior. We will explore the world of possibilities
in this chapter, while addressing details important to optimization.
Once we have established a space of candidate models, we will require
some mechanism to differentiate possible choices based on their merits, a
process known as model assessment that we will explore at length in the chapter 4: model assessment, selection, and
next chapter. We will begin the present discussion by revisiting the topic averaging, p. 67

This material will be published by Cambridge University Press as Bayesian Optimization. 45

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
modeling with gaussian processes

Figure 3.1: The importance of the prior mean function in determining sample path behavior. The models in the first two
panels differ in their mean function but share the same covariance function. Sample path behavior is identical
up to translation. The model in the third panel features the same mean function as the first panel but a different
covariance function. Samples exhibit dramatically different behavior.

of prior mean and covariance functions with an eye toward practical

utility.

3.1 the prior mean function

Recall the mean function of a Gaussian process specifies the expected
value of an arbitrary function value 𝜙 = 𝑓 (𝑥):
𝜇 (𝑥) = 𝔼[𝜙 | 𝑥].
Although this is obviously a fundamental concern, the choice of prior
mean function has received relatively little consideration in the Bayesian
optimization literature.
impact of prior mean on sample paths There are several reasons for this. To begin, it is actually the covari-
ance function rather than the mean function that largely determines the
behavior of sample paths. This should not be surprising: the mean func-
tion only affects the marginal distribution of function values, whereas
the covariance function can further modify the joint distribution of
function values. To elaborate, consider an arbitrary Gaussian process
GP (𝑓 ; 𝜇, 𝐾). Its sample paths are distributed identically to those from
the corresponding centered process 𝑓 − 𝜇, after shifting pointwise by
𝜇. Therefore the sample paths of any Gaussian process with the same
covariance function are effectively the same up to translation, and it is
the covariance function determining their behavior otherwise; see the
demonstration in figure 3.1.
impact of prior mean on posterior mean It is also important to understand the role of the prior mean func-
tion in the posterior process. Suppose we condition a Gaussian process
GP (𝑓 ; 𝜇, 𝐾) on the observation of a vector y with marginal distribution
(2.7) and cross covariance function (2.24)
𝑝 (y) = N (y; m, C); 𝜅 (𝑥) = cov[y, 𝜙 | 𝑥].
The prior mean influences the posterior process only through the poste-
rior mean (2.10):
𝜇D (𝑥) = 𝜇 (𝑥) + 𝜅 (𝑥)> C−1 (y − m).

46
 3.1. the prior mean function

Figure 3.2: The influence of the prior

mean on the posterior mean.
We show two Gaussian pro-
cess posteriors differing only
in their prior mean functions,
shown as dashed lines. In
the “interpolatory” region be-
tween the observations, the
posterior means are mostly
extrapolation interpolation extrapolation determined by the data, but
devolve to the respective
prior means when extrapolat-
We can roughly understand the behavior of the posterior mean by iden- ing outside this region.
tifying two regimes determined by the strength of correlation between a
given function value and the observations. In “interpolatory” regions,
where function values have significant correlation with one-or-more
observed value, the posterior mean is mostly determined by the data behavior in interpolatory and extrapolatory
rather than the prior mean. On the other hand, in “extrapolatory” regions, regions
where 𝜅 (𝑥) ≈ 0, the data have little influence and the posterior mean
effectively equals the prior mean. Figure 3.2 demonstrates this effect.

Constant mean function

The primary impact of the prior mean on our predictions – and on an
optimization policy informed by these predictions – is in the extrapo-
latory regime. However, extrapolation without strong prior knowledge
can be a dangerous business. As a result, in Bayesian optimization, the
prior mean is often taken to be a constant:

𝜇 (𝑥; 𝑐) ≡ 𝑐, (3.1)

in order to avoid any unwanted bias on our decisions caused by spurious

structure in the prior process. This simple choice is supported empiri- 3 g. de ath et al. (2020). What do you Mean?
cally by a study comparing optimization performance across a range of The Role of the Mean Function in Bayesian
Optimization. gecco 2020.
problems as a function of the choice of prior mean.3
When adopting a constant mean, the value of the constant 𝑐 is usually marginalizing constant prior mean
treated as a parameter to be estimated or (approximately) marginalized,
as we will discuss in the next chapter. However, we can actually do
better in some cases. Consider a parametric Gaussian process prior with model selection and averaging: §§ 4.3–4.4,
constant mean (3.1) and arbitrary covariance function: p. 73

𝑝 (𝑓 | 𝑐) = GP (𝑓 ; 𝜇 ≡ 𝑐, 𝐾),

and suppose we place a normal prior on 𝑐:

𝑝 (𝑐) = N (𝑐; 𝑎, 𝑏 2 ). (3.2)

Then we can marginalize the unknown constant mean exactly to derive

the marginal Gaussian process
∫
𝑝 (𝑓 ) = 𝑝 (𝑓 | 𝑐) 𝑝 (𝑐) d𝑐 = GP (𝑓 ; 𝜇 ≡ 𝑎, 𝐾 + 𝑏 2 ), (3.3)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 47

 modeling with gaussian processes

where the uncertainty in the mean has been absorbed into the prior
covariance function. We may now use this prior directly, avoiding any
4 Noting that 𝑐 and 𝑓 form a joint Gaussian pro- estimation of 𝑐. The unknown mean will be automatically marginalized
cess, we may perform inference as described in both the prior and posterior process, and we may additionally derive
in § 2.4, p. 26 to reveal their joint posterior.
the posterior belief over 𝑐 given data if it is of interest.4

Linear combination of basis functions

5 The basis functions can be arbitrarily complex, We may extend the above result to marginalize the weights of an arbitrary
such as the output layer of a deep neural net- linear combination of basis functions under a normal prior, making this
work:
a particularly convenient class of mean functions. Namely, consider a
j. snoek et al. (2015). Scalable Bayesian Opti- parametric mean function of the form
mization Using Deep Neural Networks. icml
2015. 𝜇 (𝑥; 𝜷) = 𝜷 >𝝍 (𝑥), (3.4)

basis functions, 𝝍 where the vector-valued function 𝝍 : X → ℝ𝑛 defines the basis functions
weight vector, 𝜷 and 𝜷 is a vector of weights.5
Now consider a parametric Gaussian process prior with a mean
function of this form (3.4) and arbitrary covariance function 𝐾. Placing
a multivariate normal prior on 𝜷,

𝑝 (𝜷) = N (𝜷; a, B), (3.5)

6 a. o’hagan (1978). Curve Fitting and Optimal and marginalizing yields the marginal prior,6, 7
Design for Prediction. Journal of the Royal Sta-
tistical Society Series B (Methodological) 40(1): 𝑝 (𝑓 ) = GP (𝑓 ; 𝑚, 𝐶),
1–42.
7 c. e. rasmussen and c. k. i. williams (2006). where
Gaussian Processes for Machine Learning. mit
Press. [§ 2.7] 𝑚(𝑥) = a>𝝍 (𝑥); 𝐶 (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) + 𝝍 (𝑥)> B𝝍 (𝑥 0). (3.6)

We may recover the constant mean case above by taking 𝜓 (𝑥) ≡ 1.

8 For an example of such modeling in physics,
where the mean function was taken to be the Other options
output of a physically informed model, see:

m. a. ziatdinov et al. (2021). Physics makes

We stress that a constant or linear mean function is by no means nec-
the difference: Bayesian optimization and ac- essary, and when a system is understood sufficiently well to suggest a
tive learning via augmented Gaussian process. plausible alternative – perhaps the output of a baseline predictive model
arXiv: 2108.10280 [physics.comp-ph]. – it should be strongly considered. However, it is hard to provide general
9 This mean function was proposed in the con- advice, as this modeling will be situation dependent.8
text of Bayesian optimization (with diagonal One option that might be a reasonable choice in some optimization
A) by
contexts is a concave quadratic mean:
j. snoek et al. (2015). Scalable Bayesian Opti-
mization Using Deep Neural Networks. icml 𝜇 (x; A, b, 𝑐) = (x − b) > A−1 (x − b) + 𝑐, (3.7)
2015,
where A ≺ 0.9 This mean encodes that values near b (according to the
who also proposed appropriate priors for A
and b. The mean was also proposed in the re- Mahalanobis distance (a.8)) are expected to be higher than those farther
lated context of Bayesian quadrature (see § 2.6, away and could reasonably model an objective function expected to be
p. 33) by “bowl-shaped” to a first approximation. The middle panel of figure 3.1
l. acerbi (2018). Variational Bayesian Monte incorporates a mean of this form; note that the maxima of sample paths
Carlo. neurips 2018. are of course not constrained to agree with that of the prior mean.

48
 3.2. the prior covariance function

3.2 the prior covariance function

The prior covariance function determines the covariance between the
function values corresponding to a pair of input locations 𝑥 and 𝑥 0:

𝐾 (𝑥, 𝑥 0) = cov[𝜙, 𝜙 0 | 𝑥, 𝑥 0]. (3.8)

The covariance function determines fundamental properties of sample sample path continuity: § 2.5, p. 28
path behavior, including continuity, differentiability, and aspects of the sample path differentiability: § 2.6, p. 30
global optima, as we have already seen. Perhaps more so than the mean existence and uniqueness of global maxima:
function, careful design of the covariance function is critical to ensure § 2.7, p. 33
fidelity in modeling. We will devote considerable discussion to this topic,
beginning with some important properties and moving on to useful
examples and mechanisms for systematically modifying and composing
multiple covariance functions together to model complex behavior.
After appropriate normalization, a covariance function 𝐾 may be
loosely interpreted as a measure of similarity between points in the
domain. Namely, given 𝑥, 𝑥 0 ∈ X, the correlation between the corre-
sponding function values is correlation between function values, 𝜌

𝐾 (𝑥, 𝑥 0)
𝜌 = corr[𝜙, 𝜙 0 | 𝑥, 𝑥 0] = √︁ , (3.9)
𝐾 (𝑥, 𝑥) 𝐾 (𝑥 ,0 𝑥 0)
and we may interpret the strength of this dependence as a measure of
similarity between the input locations. This intuition can be useful, but
some caveats are in order. To begin, note that correlation may be negative,
which might be interpreted as indicating dis-similarity as the function
values react to information with opposite sign.
Further, for a proposed covariance function 𝐾 to be admissible, it
must satisfy two global consistency properties ensuring that the collec-
tion of random variables comprising 𝑓 are able to satisfy the purported
relationships. First, we can immediately deduce from its definition (3.8)
that 𝐾 must be symmetric in its inputs. Second, the covariance function symmetry and positive semidefiniteness
must be positive semidefinite; that is, given any finite set of points x ⊂ X,
the Gram matrix 𝐾 (x, x) must have only nonnegative eigenvalues.10 10 Symmetry guarantees the eigenvalues are real.
To illustrate how positive semidefiniteness ensures statistical validity, consequences of positive semidefiniteness
note that a direct consequence is that 𝐾 (𝑥, 𝑥) = var[𝜙 | 𝑥] ≥ 0, and thus
marginal variance is always nonnegative. On a slightly less trivial level,
consider a pair of points x = (𝑥, 𝑥 0) and normalize the corresponding
Gram matrix 𝚺 = 𝐾 (x, x) to yield the correlation matrix:
 
1 𝜌
P = corr[𝝓 | x] = ,
𝜌 1

where 𝜌 is given by (3.9). For this matrix to be valid, we must have

𝜌 ∈ [−1, 1]. This happens precisely when P is positive semidefinite, as
its eigenvalues are 1 ± 𝜌. Finally, noting that P is congruent to 𝚺,11 we 11 We have
√ 𝚺 = SPS, where S is diagonal with
conclude the implied correlations are consistent if and only if 𝚺 is positive 𝑆𝑖𝑖 = Σ𝑖𝑖 .
semidefinite. With more than two points, the positive semidefiniteness
of 𝐾 ensures similar consistency at higher orders.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 49

 modeling with gaussian processes

Figure 3.3: Left: a sample from a station-

ary Gaussian process in two
dimensions. The joint distri-
bution of function values is
translation- but not rotation-
invariant, as the function
tends to vary faster in some
directions than others. Right:
a sample from an isotropic
process. The joint distribu-
tion of function values is
both translation- and rotation-
invariant.
stationary and anisotropic stationary and isotropic

Stationarity, isotropy, and Bochner’s theorem

Some covariance functions exhibit structure giving rise to certain com-
putational benefits. Namely, a covariance function 𝐾 (𝑥, 𝑥 0) that only
stationary covariance function, 𝐾 (𝑥 − 𝑥 0 ) depends on the difference 𝑥 − 𝑥 0 is called stationary.12 When convenient,
we will abuse notation and write a stationary covariance function in
12 Of course this definition requires 𝑥 − 𝑥 0 to terms of a single input, writing 𝐾 (𝑥 − 𝑥 0) for 𝐾 (𝑥, 𝑥 0) = 𝐾 (𝑥 − 𝑥 ,0 0). If a
be well defined. This is trivial in Euclidean gp has a stationary covariance function and constant mean function (3.1),
spaces; a fairly general treatment for more
exotic spaces would assume an abelian group
then the process itself is also called stationary. A consequence of station-
structure on X with binary operation + and arity is that the distribution of any set of function values is invariant
inverse − and define 𝑥 − 𝑥 0 = 𝑥 + (−𝑥 0 ). under translation; that is, the function “acts the same” everywhere from
a statistical viewpoint. The left panel of figure 3.3 shows a sample from
a 2𝑑 stationary gp, demonstrating this translation-invariant behavior.
Stationarity is a convenient assumption when modeling, as defining
the local behavior around a single point suffices to specify the global
behavior of an entire function. Many common covariance functions
have this property as a result. However, this may not always be a valid
assumption in the context of optimization, as an objective function may
for example exhibit markedly different behavior near the optimum than
elsewhere. We will shortly see some general approaches for addressing
nonstationarity when appropriate.
If X ⊂ ℝ𝑛, a covariance function 𝐾 (𝑥, 𝑥 0) only depending on the
isotropic covariance function, 𝐾 (𝑑) Euclidean distance 𝑑 = |𝑥 −𝑥 0 | is called isotropic. Again, when convenient,
we will notate such a covariance with 𝐾 (𝑑). Isotropy is a more restrictive
assumption than stationarity – indeed it trivially implies stationarity –
as it implies the covariance is invariant to both translation and rotation,
and thus the function has identical behavior in every direction from
13 s. bochner (1933). Monotone Funktionen, every point. An example sample from a 2𝑑 isotropic gp is shown in the
Stieltjessche Integrale und harmonische Anal- right panel of figure 3.3. Many of the standard covariance functions
yse. Mathematische Annalen 108:378–410. we will define shortly will be isotropic on first definition, but we will
14 We do not quote the most general version of again develop generic mechanisms to modify them in order to induce
the theorem here; the result can be extended anisotropic behavior when desired.
to complex-valued covariance functions on ar-
bitrary locally compact abelian groups if nec- bochner’s theorem is an landmark result characterizing stationary
essary. It is remarkably universal. covariance functions in terms of their Fourier transforms:13, 14

50
 3.3. notable covariance functions

Theorem (bochner, 1933). A continuous function 𝐾 : ℝ𝑛 → ℝ is positive

semidefinite (that is, represents a stationary covariance function) if and
only if we have ∫
𝐾 (x) = exp(2𝜋𝑖x>𝝃 ) d𝜈,

where 𝜈 is a finite, positive Borel measure on ℝ𝑛. Further, this measure

is symmetric around the origin; that is, 𝜈 (𝐴) = 𝜈 (−𝐴) for any Borel set
𝐴 ⊂ ℝ𝑛, where −𝐴 is the “negation” of 𝐴: −𝐴 = {−𝑎 | 𝑎 ∈ 𝐴}.
To summarize, bochner’s theorem states that the Fourier transform of
any stationary covariance function on ℝ𝑛 is proportional to a probability
measure and vice versa; the constant of proportionality is 𝐾 (0). The
measure 𝜈 corresponding to 𝐾 is called the spectral measure of 𝐾. When spectral measure, 𝜈
a corresponding density function 𝜅 exists, it is called the spectral density spectral density, 𝜅
of 𝐾 and forms a Fourier pair with 𝐾:
∫ ∫
𝐾 (x) = exp(2𝜋𝑖x>𝝃 ) 𝜅 (𝝃 ) d𝝃 ; 𝜅 (𝝃 ) = exp(−2𝜋𝑖x>𝝃 ) 𝐾 (x) dx.
(3.10)
The symmetry of the spectral measure implies a similar symmetry in symmetry of spectral density
the spectral density: 𝜅 (𝝃 ) = 𝜅 (−𝝃 ) for all 𝝃 ∈ ℝ𝑛.
bochner’s theorem is surprisingly useful in practice, allowing us to
approximate an arbitrary stationary covariance function by approximat-
ing (e.g., by modeling or sampling from) its spectral density. This is the
basis of the spectral mixture covariance described in the next section, as
well as the sparse spectrum approximation scheme, which facilitates the sparse spectrum approximation: § 8.7, p. 178
computation of some Bayesian optimization policies.

3.3 notable covariance functions

It can be difficult to define a new covariance function for a given sce-
nario de novo, as the positive-semidefinite criterion can be nontrivial to
guarantee for what might otherwise be an intuitive notion of similarity.
In practice, it is common to instead construct covariance functions by
combining and transforming established “building blocks” modeling var-
ious atomic behaviors while following rules guaranteeing the result will 15 For a more complete survey, see
be valid. We describe several useful examples below.15 c. e. rasmussen and c. k. i. williams (2006).
Our presentation will depart from most in that several of the covari- Gaussian Processes for Machine Learning. mit
ance functions below will initially be defined without parameters that Press. [chapter 4]
some readers may be expecting. We will shortly demonstrate how cou-
pling these covariance functions with particular transformations of the
function domain and output gives rise to common covariance function
parameters such as characteristic length scales and output scales.

The Matérn family

If there is one class of covariance functions to be familiar with, it is
the Matérn family. This is a versatile family of covariance functions
for modeling isotropic behavior on Euclidean domains X ⊂ ℝ𝑛 of any

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 51

 modeling with gaussian processes

𝜈 = 1/2, (3.11) 𝜈 = 3/2, (3.13) 𝜈 = 5/2, (3.14)

Figure 3.4: Samples from centered Gaussian processes with the Matérn covariance function with different values of the
smoothness parameter 𝜈. Sample paths with 𝜈 = 1/2 are continuous but not differentiable; incrementing this
parameter by one unit increases the number of continuous derivatives by one.

desired degree of smoothness, in terms of the differentiability of sample

sample path differentiability: § 2.6, p. 30 paths. The Matérn covariance 𝐾M(𝜈) depends on a parameter 𝜈 ∈ ℝ>0
determining this smoothness; sample paths from a centered Gaussian
process with this covariance are d𝜈e − 1 times continuously differentiable.
16 In theoretical contexts, general values for the In practice 𝜈 is almost always taken to be a half-integer,16 in which case
smoothness parameter 𝜈 ∈ 𝑅 >0 are consid- the expression for the covariance assumes a simple form as a function
ered, but lead to unwieldy expressions (10.12).
of the Euclidean distance 𝑑 = |𝑥 − 𝑥 0 |.
To begin with the extremes, the case 𝜈 = 1/2 yields the so-called
exponential covariance function exponential covariance:

𝐾M1/2 (𝑥, 𝑥 0) = exp(−𝑑). (3.11)

Sample paths from a centered Gaussian process with exponential co-

variance are continuous but nowhere differentiable, which is perhaps
too rough to be interesting in most optimization contexts. However,
this covariance is often encountered in historical literature. In the one-
Ornstein–Uhlenbeck (ou) process dimensional case X ⊂ ℝ, a Gaussian process with this covariance is
known as a Ornstein–Uhlenbeck (ou) process and satisfies a continuous-
time Markov property that renders its posterior moments particularly
convenient.
Taking the limit of increasing smoothness 𝜈 → ∞ yields the squared
exponential covariance from the previous chapter:


𝐾se (𝑥, 𝑥 0) = exp − 12 𝑑 2 . (3.12)

Samples from a centered Gaussian process We will refer to the Matérn and the limiting case of the squared expo-
with squared exponential covariance 𝐾se . nential covariance functions together as the Matérn family. The squared
exponential covariance is without a doubt the most prevalent covari-
ance function in the statistical and machine learning literature. However,
it may not always be a good choice in practice. Sample paths from a
centered Gaussian process with squared exponential covariance are in-
17 m. l. stein (1999). Interpolation of Spatial Data: finitely differentiable, which has been ridiculed as an absurd assumption
Some Theory for Kriging. Springer–Verlag. for most physical processes.17 stein does not mince words on this, start-
[§ 1.7] ing off a three-sentence “summary of practical suggestions” with “use
the Matérn model” and devoting significant effort to discouraging the
use of the squared exponential in the context of geostatistics.

52
 3.3. notable covariance functions

Between these extremes are the cases 𝜈 = 3/2 and 𝜈 = 5/2, which
respectively model once- and twice-differentiable functions:
√
√

𝐾M3/2 (𝑥, 𝑥 0) = 1 + 3𝑑 exp − 3𝑑 ; (3.13)
√
√

𝐾M5/2 (𝑥, 𝑥 0) = 1 + 5𝑑 + 53 𝑑 2 exp − 5𝑑 . (3.14)
Figure 3.4 illustrates samples from centered Gaussian processes with
different values of the smoothness parameters 𝜈. The 𝜈 = 5/2 case in 18 j. snoek et al. (2012). Practical Bayesian Op-
particular has been singled out as a prudent off-the-shelf choice for timization of Machine Learning Algorithms.
neurips 2012.
Bayesian optimization when no better alternative is obvious.18

The spectral mixture covariance

Covariance functions in the Matérn family express fairly simple corre-
lation structure, with the covariance dropping monotonically to zero 𝐾 (𝑥, 𝑥 0 ) 𝐾se
as the distance 𝑑 = |𝑥 − 𝑥 0 | increases. All differences in sample path 𝐾M3/2
behavior such as differentiability, etc. are expressed entirely through 𝐾M1/2
nuances in the tail behavior of the covariance functions; see the figure
in the margin.
The Fourier transforms of these covariances are also broadly com-
parable: all are proportional to unimodal distributions centered on the 𝑑
origin. However, bochner’s theorem indicates that there is a vast world
of stationary covariance functions indexed by the entire space of sym- Some members of the Matérn family and the
metric spectral measures, which may have considerably more complex squared exponential covariance as a function
of the distance between inputs. All decay to
structure. Several authors have sought to exploit this characterization to zero correlation as distance increases.
build stationary covariance functions with virtually unlimited flexibility.
A notable contribution in this direction is the spectral mixture covari-
ance function proposed by wilson and adams.19 The idea is simple but 19 a. g. wilson and r. p. adams (2013). Gaussian
powerful: we parameterize a space of stationary covariance functions Process Kernels for Pattern Discovery and Ex-
trapolation. icml 2013.
by some suitable family of mixture distributions in the Fourier domain
representing their spectral density. The parameters of this spectral mix-
ture distribution specify a covariance function via the correspondence
in (3.10), and we can make the resulting family as rich as desired by
adjusting the number of components in the mixture. wilson and adams
proposed Gaussian mixtures for the spectral density, which are universal
approximators and have a convenient Fourier transform. We define a
Gaussian mixture spectral density 𝜅 as
∑︁  
𝑘 (𝝃 ) = 𝑤𝑖 N (𝝃 ; 𝝁𝑖 , 𝚺𝑖 ); 𝜅 (𝝃 ) = 21 𝑘 (𝝃 ) + 𝑘 (−𝝃 ) ,
𝑖

where the indirect construction via 𝑘 ensures the required symmetry.

Note that the weights {𝑤𝑖 } must be positive but need not sum to unity.
Taking the inverse Fourier transform (3.10), the corresponding covariance
function is


𝐾sm x, x0; {𝑤𝑖 }, {𝝁𝑖 }, {𝚺𝑖 } = Samples from centered Gaussian processes
∑︁

with two realizations of a Gaussian spectral
𝑤𝑖 exp −2𝜋 2 (x − x0)> 𝚺𝑖 (x − x0) cos 2𝜋 (x − x0)>𝝁𝑖 . (3.15) mixture covariance function, offering a
𝑖 glimpse into the flexibility of this class.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 53

 modeling with gaussian processes

Inspecting this expression, we can see that every covariance function

induced by a Gaussian mixture spectral density is infinitely differentiable,
and one might object to this choice on the grounds of overly smooth sam-
ple paths. This can be mitigated by using enough mixture components
to induce sufficiently complex structure in the covariance (on the order
of ∼5 is common). Another option would be to use a different family of
spectral distributions; for example, a mixture of Cauchy distributions
would induce a family of continuous but nondifferentiable covariance
functions analogous to the exponential covariance (3.11), but this idea
has not been explored.

Linear covariance function

Another useful covariance function arises from a Bayesian realization
20 Independence is usual but not necessary; an of linear regression. Let the domain be Euclidean, X ⊂ ℝ,𝑛 and consider
arbitrary joint prior would add a term of 2b> x the model
to (3.16), where b = cov[𝜷, 𝛽 ].
𝑓 (x) = 𝛽 + 𝜷 >x,
where we have abused notation slightly to distinguish the constant term
linear basis functions: § 3.1, p. 48 from the remaining coefficients. Following our discussion on linear basis
functions, if we take independent20 normal priors on 𝛽 and 𝜷:

𝑝 (𝛽) = N (𝛽; 𝑎, 𝑏 2 ); 𝑝 (𝜷) = N (𝜷; a, B),

we arrive at the so-called linear covariance:

𝐾lin (x, x0; 𝑏, B) = 𝑏 2 + x>Bx. (3.16)

Although this covariance is unlikely to be of any direct use in Bayesian

Samples from a centered Gaussian process
with linear covariance 𝐾lin . optimization (linear programming is much simpler!), it can be a useful
component of more complex composite covariance structures.

3.4 modifying and combining covariance functions

With the notable exception of the spectral mixture covariance, which
can approximate any stationary covariance function, several of the co-
variances introduced in the last section are still too rigid to be useful.
In particular, consider any of the Matérn family (3.11–3.14). Each
of these covariances encodes several explicit and possibly dubious as-
sumptions about the function of interest. To begin, each prescribes unit
variance for every function value:

var[𝜙 | 𝑥] = 𝐾 (𝑥, 𝑥) = 1, (3.17)

which is an arbitrary, possibly inappropriate choice of scale. Further, each

21 Although an important concept, there is no of these covariance functions fixes an isotropic characteristic length scale
clear-cut definition of characteristic length of correlation of approximately one unit:21 at a separation of |𝑥 − 𝑥 0 | = 1,
scale. It is simply a convenient separation dis-
tance for which correlation remains apprecia- the correlation between the corresponding function values drops to
ble, but beyond which correlation begins to roughly
noticeably decay. corr[𝜙, 𝜙 0 | 𝑥, 𝑥 0] ≈ 0.5, (3.18)

54
 3.4. modifying and combining covariance functions

prior mean prior 95% credible interval samples

Figure 3.5: Scaling a stationary covari-

ance by a nonconstant func-
tion (here, a smooth bump
function of compact support)
to yield a nonstationary co-
variance.
scaling function, 𝑎
1

and this correlation continues to drop effectively to zero at a separation

of approximately five units. Again, this choice of scale is arbitrary, and
the assumption of isotropy is particularly restrictive.
In general, a Gaussian process encodes strong assumptions regarding
the joint distribution of function values (2.5), which may not be com-
patible with a given function “out of the box.” However, we can often
improve model fit by appropriate transformations of the objective. In fact,
linear transformations of function inputs and outputs are almost uni-
versally considered, although only implicitly by introducing parameters
conveying the effects of these transformations. We will show how both
linear and nonlinear transformations of function input and output lead
to more expressive models and give rise to common model parameters.

Scaling function outputs

We first address the issue of scale in function output (3.17) by considering
the statistical effects of arbitrary scaling. Consider a random function
𝑓 : X → ℝ with covariance function 𝐾 and let 𝑎 : X → ℝ be a known
scaling function.22 Then the pointwise product 𝑎𝑓 : 𝑥 ↦→ 𝑎(𝑥) 𝑓 (𝑥) has 22 For this result 𝑓 need not have a gp distribu-
covariance function tion.

cov[𝑎𝑓 | 𝑎] = 𝑎(𝑥)𝐾 (𝑥, 𝑥 0)𝑎(𝑥 0), (3.19)

by the bilinearity of covariance. If the scaling function is constant, 𝑎 ≡ 𝜆,
then we have
cov[𝜆𝑓 | 𝜆] = 𝜆 2𝐾 . (3.20)
This simple result allows us to extend a “base” covariance 𝐾 with fixed
scale, as in (3.17), to a parametric family with arbitrary scale:
𝐾 0 (𝑥, 𝑥 0; 𝜆) = 𝜆 2 𝐾 (𝑥, 𝑥 0).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 55

modeling with gaussian processes

𝐾
In this context the parameter 𝜆 is known as an output scale, or when
the base covariance is stationary with 𝐾 (𝑥, 𝑥) = 1, the signal variance,
as it determines the variance of any function value: var[𝜙 | 𝑥, 𝜆] = 𝜆 2.
The illustration in the margin shows the effect of scaling the squared
exponential covariance function by a series of increasing output scales.
We can also of course consider nonlinear transformations of the
𝑑
function output as well. This can be useful for modeling constraints –
such as nonegativity or boundedness – that are not compatible with
The squared exponential covariance 𝐾se the Gaussian assumption. However, a nonlinear transformation of a
scaled by a range of output scales 𝜆 (3.20). Gaussian process is no longer Gaussian, so it often more convenient to
model the transformed function after “removing the constraint.”
We may use the general form of this scaling result (3.19) to transform
a stationary covariance into a nonstationary one, as any nonconstant
scaling is sufficient to break translation invariance. We show an example
of such a transformation in figure 3.5, where we have scaled a stationary
covariance by a bump function to create a prior on smooth functions
with compact support.

Transforming the domain and length scale parameters

We now address the issue of the scaling of correlation as a function of
Sample paths from centered gps with smaller distance (3.18) by introducing a powerful tool: transforming the domain
(top) and larger (bottom) output scales.
of the function of interest into a more convenient space for modeling.
Namely, suppose we wish to reason about a function 𝑓 : X → ℝ,
𝐾
and let 𝑔 : X → Z be a map from the domain to some arbitrary space
Z, which might also be X . If 𝐾Z is a covariance function on Z, then the
composition

𝐾X (𝑥, 𝑥 0) = 𝐾Z 𝑔(𝑥), 𝑔(𝑥 0) (3.21)
is trivially a covariance function on X. This allows us to define a covari-
ance for 𝑓 indirectly by jointly designing a map 𝑔 to another space and
𝑑 a corresponding covariance 𝐾Z (and mean 𝜇Z ) on that space. This ap-
proach offers a lot of flexibility, as we are free to design these components
The squared exponential covariance 𝐾se as we see fit to impose any desired structure.
dilated by a range of length scales ℓ (3.22). We will spend some time exploring this idea, beginning with the
relatively simple but immensely useful case of combining a linear trans-
formation on a Euclidean domain X ⊂ ℝ𝑛 with an isotropic covariance
on the output. Perhaps the simplest example is the dilation x ↦→ x/ℓ,
which simply scales distance by ℓ −1 . Incorporating this transformation
into an isotropic base covariance 𝐾 (𝑑) on X yields a parametric family
of dilated versions:
𝐾 0 (𝑥, 𝑥 0; ℓ) = 𝐾 (𝑑/ℓ). (3.22)
If the base covariance has a characteristic length scale of one unit, the
length scale of the dilated version will be ℓ; for this reason, this parameter
is simply called the characteristic length scale of the parameterized
Sample paths from centered gps with shorter covariance (3.22). Adjusting the length scale allows us to model functions
(top) and longer (bottom) characteristic with a range of “wiggliness,” where shorter length scale implies more
length scales.
wiggly behavior; see the margin for examples.

56
 3.4. modifying and combining covariance functions

Figure 3.6: Left: a sample from a cen-

tered Gaussian process in two
dimensions with isotropic
squared exponential covari-
ance. Right: a sample from
a centered Gaussian process
with an ard squared exponen-
tial covariance. The length of
the lines on each axis are pro-
portional to the length scale
along that axis.

Taking this one step further, we may consider dilating each axis by a
separate factor:

𝑥𝑖 ↦→ 𝑥𝑖 /ℓ𝑖 ; x ↦→ [diag ℓ] −1 x, (3.23)

which induces the weighted Euclidean distance

√︄
∑︁ (𝑥 − 𝑥 0) 2
𝑑ℓ = 𝑖 𝑖
. (3.24)
𝑖
ℓ𝑖

Geometrically, the effect of this map is to transform surfaces of equal dis-

tance around each point – which represent curves of constant covariance
for an isotropic covariance – from spheres into axis-aligned ellipsoids;
see the figure in the margin. Incorporating into an isotropic base covari-
ance 𝐾 (𝑑) produces a parametric family of anisotropic covariances with Possible surfaces of equal covariance with the
different characteristic length scales along each axis, corresponding to center when combining separate dilation of
each axis with an isotropic covariance.
the parameters ℓ:
𝐾 0 (𝑥, 𝑥 0; ℓ) = 𝐾 (𝑑 ℓ ). (3.25)
When the length scale parameters are inferred from data, this construc-
tion is known as automatic relevance determination (ard). The motivation automatic relevance determination, ard
for the name is that if the function has only weak dependence on some
mostly irrelevant dimension of the input, we could hope to infer a very
long length scale for that dimension. The contribution to the weighted
distance (3.24) for that dimension would then be effectively nullified,
and the resulting covariance would effectively “ignore” that dimension.
Figure 3.6 shows samples from 2𝑑 centered Gaussian processes, com-
paring behavior with an isotropic covariance and an ard modified ver-
sion that contracts the horizontal and expands the vertical axis (see
curves of constant covariance in the margin). The result is anisotropic
behavior with a longer characteristic length scale in the vertical direction
Surfaces of equal covariance with the center
than in the horizontal direction, but with the behavior of local features for the examples in figure 3.6: the isotropic
remaining aligned with the axes overall. covariance in the left panel (red), and the ard
Finally, we may also consider an arbitrary linear transformation covariance in the right panel (blue).
𝑔 : x ↦→ Ax, which induces the Mahalanobis distance (a.8)

𝑑 A = |Ax − Ax0 |.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 57

 modeling with gaussian processes

As before, we may incorporate this map into an isotropic base covariance

𝐾 to realize a family of anisotropic covariance functions:
𝐾 0 (𝑥, 𝑥 0; A) = 𝐾 (𝑑 A ). (3.26)
Geometrically, an arbitrary linear map can transform surfaces of constant
covariance from spheres into arbitrary ellipsoids; see the figure in the
margin. The sample from the left-hand side of figure 3.3 was generated by
composing an isotropic covariance with a map inducing both anisotropic
scaling and rotation. The effect of the underlying transformation can be
seen in the shapes of local features, which are not aligned with the axes.
Due to the inherent number of parameters required to specify a
general transformation, this construction is perhaps most useful when
the map is to a much lower-dimensional space: ℝ𝑛 → ℝ,𝑘 𝑘  𝑛. This
Possible surfaces of equal covariance with the
has been promoted as one strategy for modeling functions on high-
center when combining an arbitrary linear
transformation with an isotropic covariance. dimensional domains suspected of having hidden low-dimensional struc-
ture – if this low-dimensional structure is along a linear subspace of the
domain, we could capture it by an appropriately designed projection A.
high-dimensional domains: § 3.5, p. 61 We will discuss this idea further in the next section.

Nonlinear warping
When using a covariance function with an inherent length scale, such as
a Matérn or squared exponential covariance, some linear transformation
of the domain is almost always considered, whether it be simple dilation
(3.22), anisotropic scaling (3.25), or a general transformation (3.26). How-
23 We did see a periodic gp in the previous chap- ever, nonlinear transformations can also be useful for imposing structure
ter (2.30); however, that model only had sup-
on the domain, a process commonly referred to as warping.
port on perfectly sinusoidal functions.
To provide an example that may not often be useful in optimization
24 d. j. c. mackay (1998). Introduction to Gaus-
but is illustrative nonetheless, suppose we wish to model a function
sian Processes. Neural Networks and Machine
Learning. Vol. 168. Springer–Verlag. [§ 5.2] 𝑓 : ℝ → ℝ that we believe to be smooth and periodic with period 𝑝.
25 The covariance on the circle is usually
None of the covariance functions introduced thus far would be able to
inherited from a covariance on ℝ2. The result induce the periodic correlations that this assumption would entail.23 A
of composing with the squared exponential construction due to mackay is to compose a map onto a circle of radius
covariance in particular is often called “the”
periodic covariance, but we stress that any
𝑟 = 𝑝/(2𝜋):24  
other covariance on ℝ2 could be used instead.
𝑟 cos 𝑥
𝑥 ↦→ (3.27)
𝑟 sin 𝑥
with a covariance function on that space reflecting any desired properties
of 𝑓.25 As this map identifies points separated by any multiple of the
period, the corresponding function values are perfectly correlated, as
desired. A sample from a Gaussian process employing this construction
with a Matérn covariance after warping is shown in the margin.
A compelling use of warping is to build nonstationary models by
A sample path of a centered gp with Matérn
covariance with 𝜈 = 5/2 (3.14) after applying
composing a nonlinear map with a stationary covariance, an idea snoek
the periodic warping function (3.27). et al. explored in the context of Bayesian optimization.26 Many objec-
tive functions exhibit different behavior depending on the proximity to
26 j. snoek et al. (2014). Input Warping for Bayes- the optimum, suggesting that nonstationary models may sometimes be
ian Optimization of Non-Sationary Functions. worth exploring. snoek et al. proposed a flexible family of warping func-
icml 2014. tions for optimization problems with box-bounded constraints, where

58
 3.4. modifying and combining covariance functions

prior mean prior 95% credible interval samples

1 Figure 3.7: Am example of the beta warping method proposed

by snoek et al. We show three samples of the
stationary Gaussian process prior from figure 2.1
(above) after applying a nonlinear warping through
a beta cdf (3.28) with (𝛼, 𝛽) = (4, 4) (right). The
length scale is compressed in the center of the do-
main and expanded near the boundary.
0
0 1

we may take the domain to be the unit cube by scaling and translating as
necessary: X = [0, 1] 𝑛. The idea is to warp each coordinate of the input
via the cumulative distribution function of a beta distribution: 1

𝑥𝑖 ↦→ 𝐼 (𝑥𝑖 ; 𝛼𝑖 , 𝛽𝑖 ), (3.28)

where (𝛼𝑖 , 𝛽𝑖 ) are shape parameters and 𝐼 is the regularized beta func-
tion. This represents a monotonic bijection on the unit interval that can
assume several shapes; see the marginal figure for examples. The map
may contract portions of the domain and expand others, effectively de- 0
creasing and increasing the length scale in those regions. Finally, taking 0 1
𝛼 = 𝛽 = 1 recovers the identity map, allowing us to degrade gracefully Some examples of beta cdf warping functions
to the unwarped case if desired. (3.28).
In figure 3.7 we combine a beta warping on a one-dimensional domain
with a stationary covariance on the output. The chosen warping shortens
the length scale near the center of the domain and extends it near the
boundary, which might be reasonable for an objective function expected
to exhibit the most “interesting” behavior on the interior of its domain.
A recent innovation is to use sophisticated artificial neural networks
as warping maps for modeling functions of high-dimensional data with
complex structure. Notable examples of this approach include the fami-
lies of manifold Gaussian processes introduced by calandra et al.27 and 27 r. calandra et al. (2016). Manifold Gaussian
deep kernels introduced contemporaneously by wilson et al.28 Here the Processes for Regression. ijcnn 2016.
warping function was taken to be an arbitrary neural network, the output 28 a. g. wilson et al. (2016). Deep Kernel Learn-
layer of which was fed into a suitable stationary covariance function. ing. aistats 2016.
This gives a highly parameterized covariance function where the pa-
rameters of the base covariance and the neural map become parameters
of the resulting model. In the context of Bayesian optimization, this
can be especially useful when there is sufficient data to learn a useful neural representation learning: § 8.11, p.198
representation of the domain via unsupervised methods.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 59

 modeling with gaussian processes

𝐾1 𝐾2 𝐾1 + 𝐾2
Figure 3.8: Samples from centered Gaussian processes with different covariance functions: (left) a squared exponential
covariance, (middle) a squared exponential covariance with smaller output scale and shorter length scale, and
(right) the sum of the two. Samples from the process with the sum covariance show smooth variation on two
different scales.

Combining covariance functions

In addition to modifying covariance functions via scaling the output
and/or transforming the domain, we may also combine multiple co-
variance functions together to model functions influenced by multiple
random processes.
Let 𝑓, 𝑔 : X → ℝ be two centered, independent (not necessarily Gaus-
sian) random functions with covariance functions 𝐾 𝑓 and 𝐾𝑔 , respectively.
By the properties of covariance, the sum and pointwise product of these
functions have covariance functions with the same structure:

cov[𝑓 + 𝑔] = 𝐾 𝑓 + 𝐾𝑔 ; cov[𝑓 𝑔] = 𝐾 𝑓 𝐾𝑔 , (3.29)

and thus covariance functions are closed under addition and pointwise
29 The assumption of the processes being cen- multiplication.29 Combining this result with (3.20), we have that any
tered is needed for the product result only; polynomial of covariance functions with nonnegative coefficients forms
otherwise, there would be additional terms
involving scaled versions of each individual
a valid covariance. This enables us to construct infinite families of in-
covariance as in (3.19). The sum result does creasingly complex covariance functions from simple components.
not depend on any assumptions regarding the We may use a sum of covariance functions to model a function with
mean functions. independent additive contributions, such as random behavior on several
length scales. Precisely such a construction is illustrated in figure 3.8. If
the covariance functions are nonnegative and have roughly the same
scale, the effect of addition is roughly one of logical disjunction: the sum
will assume nontrivial values whenever any one of its constituents does.
Meanwhile, a product of covariance functions can loosely be in-
terpreted in terms of logical conjunction, with function values having
appreciable covariance only when every individual covariance function
does. A prototypical example of this effect is a covariance function mod-
eling functions that are “almost periodic,” formed by the product of a
bump-shaped isotropic covariance function such as a squared exponen-
tial (3.12) with a warped version modeling perfectly periodic functions
(3.27). The former moderates the influence of the latter by driving the cor-
relation between function values to zero for inputs that are sufficiently
A sample from a centered Gaussian process separated, regardless of their positions in the periodic cycle. We show a
with an “almost periodic” covariance function.
sample from such a covariance in the margin, where the length scale of
the modulation term is three times the period.

60
 3.5. modeling functions on high-dimensional domains

3.5 modeling functions on high-dimensional domains

Optimization on a high-dimensional domain can be challenging, as we
can succumb to the curse of dimensionality if we are not careful. As an curse of dimensionality
example, consider optimizing an objective function on the unit cube
[0, 1] 𝑛. Suppose we model this function with an isotropic covariance
from the Matérn family, taking the length scale to be ℓ = 1/10 so that
ten length scales span the domain along each axis.30 This choice implies 30 This is far from excessive: the domain for the
that function values on the corners of the domain would be effectively marginal sampling examples in this chapter
spans 15 length scales and there’s just enough
independent, as exp(−10) < 10−4 (3.11) and exp(−50) is smaller still room for interesting behavior to emerge.
(3.12). If we were to demand even a modicum of confidence in these
regions at termination, say by having a measurement within one length
scale of every corner, we would need 2𝑛 observations! This exponential
growth in the number of observations required to cover the domain is
the tyrannical curse of dimensionality.
However, compelling objectives do not tend to have so many degrees 31 j. bergstra and y. bengio (2012). Random
of freedom; if they did, we should perhaps give up on the idea of global Search for Hyper-Parameter Optimization.
Journal of Machine Learning Research 13:281–
optimization altogether. Rather, many authors have noted a tendency 305.
toward low intrinsic dimensionality in real-world problems: that is, most
32 c. li et al. (2018a). Measuring the Intrinsic Di-
of the variation in the objective is confined to a low-dimensional sub- mension of Objective Landscapes. iclr 2018.
space of the domain. This phenomenon has been noted for example in
hyperparameter optimization31 and optimizing the parameters of neural
networks.32 levina and bickel suggested that “hidden” low-dimensional
structure is actually a universal requirement for success on any task:33

There is a consensus in the high-dimensional data analysis commu- 33 e. levina and p. j. bickel (2004). Maximum
nity that the only reason any methods work in very high dimensions Likelihood Estimation of Intrinsic Dimension.
neurips 2004.
is that, in fact, the data are not truly high dimensional.

The global optimization community shares a similar consensus: typical

high-dimensional objectives are not “truly” high dimensional. This intu-
ition presents us with an opportunity: if we could only identify inherent
low-dimensional structure during optimization, we could sidestep the
curse of dimensionality by restricting our search accordingly.
Several strategies are available for capturing low intrinsic dimension
with Gaussian process models. The general approach closely follows our
discussion from the previous section: we identify some appropriate map-
ping from the high-dimensional domain to a lower-dimensional space,
then model the objective function after composing with this embedding
(3.21). This is one realization of the general class of manifold Gaussian
processes,34 where the sought-after manifold is low dimensional. Adopt- 34 r. calandra et al. (2016). Manifold Gaussian
ing this approach then raises the issue of identifying useful families of Processes for Regression. ijcnn 2016.
mappings that can suitably reduce dimension while preserving enough
structure of the objective to keep optimization feasible.

Neural embeddings
Given the success of deep learning in designing feature representations
for complex, high-dimensional objects, neural embeddings – as used in

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 61

 modeling with gaussian processes

Figure 3.9: An objective function on

a two-dimensional domain
(left) with intrinsic dimension
1. The entire variation of the
objective is determined on the
one-dimensional linear sub-
space Z corresponding to the
diagonal black line, which we
can model in its inherent di- Z
mension (right).
Z

35 a. g. wilson et al. (2016). Deep Kernel Learn- the family of deep kernels35 – present a tantalizing option. Neural embed-
ing. aistats 2016. dings have shown some success in Bayesian optimization, where they
can facilitate optimization over complex structured objects by providing
neural representation learning: § 8.11, p.198 a nice continuous latent space to work in.
snoek et al. demonstrated excellent performance on hyperparameter
tuning tasks by interpreting the output layer of a deep neural network as
a set of custom nonlinear basis functions for Bayesian linear regression,
36 j. snoek et al. (2015). Scalable Bayesian Opti- as in (3.6).36 An advantage of this particular construction is that Gaussian
mization Using Deep Neural Networks. icml process inference and prediction is accelerated dramatically by adopting
2015.
the linear covariance (3.6) – the cost of inference scales linearly with the
cost of Gaussian process inference: § 9.1, p. 201 number of observations, rather than cubically as in the general case.

Linear embeddings
Another line of attack is to search for a low-dimensional linear subspace
of the domain encompassing the relevant variation in inputs and model
the function after projection onto that space. For an objective 𝑓 on a
high-dimensional domain X ⊂ ℝ,𝑛 we consider models of the form

𝑓 (x) = 𝑔(Ax); A ∈ ℝ,𝑘×𝑛 (3.30)

where 𝑔 : ℝ𝑘 → ℝ is a (𝑘  𝑛)-dimensional surrogate for 𝑓.

The simplest such approach is automatic relevance determination
37 j. bergstra and y. bengio (2012). Random (3.25), where we learn separate length scales along each dimension.37
Search for Hyper-Parameter Optimization. Although the corresponding linear transformation (3.23) does not reduce
Journal of Machine Learning Research 13:281– dimension, axes with sufficiently long length scales are effectively elimi-
305.
nated, as they do not have strong influence on the covariance. This can
be effective when some dimensions are likely to be irrelevant, but limits
us to axis-aligned subspaces only.
A more flexible option is to consider arbitrary linear transforma-
38 f. vivarelli and c. k. i. williams (1998). Dis-
covering hidden features with Gaussian pro-
tions in the model (3.26, 3.30), an idea that has seen significant attention
cess regression. neurips 1998. for Gaussian process modeling in general38 and for Bayesian optimiza-
39 z. wang et al. (2016b). Bayesian Optimization
tion in particular.39 Figure 3.9 illustrates a simple example where a one-
in a Billion Dimensions via Random Embed- dimensional objective function is embedded in two dimensions in a non-
dings. Journal of Artificial Intelligence Research axis-aligned manner. Both axes would appear important for explaining
55:361–387 the function when using ard, but a one-dimensional subspace suffices

62
 3.5. modeling functions on high-dimensional domains

Figure 3.10: A sample from a gp in two dimensions with the

𝑥 2∗ decomposition 𝑓 (x) = 𝑔1 (𝑥 1 ) + 𝑔2 (𝑥 2 ). Here a
nominally two-dimensional function is actually
the sum of two one-dimensional components de-
fined along each axis with no interaction. The
maximum of the function is achieved at the point
corresponding to the maxima of the individual
components.

𝑥 1∗

if chosen carefully. This approach offers considerably more modeling

flexibility than ard at the expense of a 𝑘-fold increase in the number
of parameters that must be specified. However, several algorithms have
been proposed for efficiently identifying a suitable map A,40, 41 and wang 40 j. djolonga et al. (2013). High-Dimensional
et al. demonstrated success in optimizing objectives in extremely high Gaussian Process Bandits. neurips 2013.
dimension by simply searching along a random low-dimensional sub- 41 r. garnett et al. (2014). Active Learning of
space. The authors also provided theoretical guarantees regarding the Linear Embeddings for Gaussian Processes.
uai 2014.
recoverability of the global optimum with this approach, assuming the
hypothesis of low intrinsic dimensionality holds.
If more flexibility is desired, we may represent an objective function
as a sum of contributions on multiple relevant linear subspaces:
∑︁
𝑓 (x) = 𝑔𝑖 (A𝑖 x). (3.31)
𝑖

This decomposition is similar in spirit to the classical family of gener-

alized additive models,42 where the linear maps can be arbitrary and of 42 t. hastie and r. tibshirani (1986). General-
variable dimension. If we assume the additive components in (3.31) are ized Additive Models. Statistical Science 1(3):
297–318.
independent, each with Gaussian process prior GP (𝑔𝑖 ; 𝜇𝑖 , 𝐾𝑖 ), then the
resulting model for 𝑓 is a Gaussian process with additive moments (3.29):
∑︁ ∑︁
𝜇 (x) = 𝜇𝑖 (A𝑖 x); 𝐾 (x, x0) = 𝐾𝑖 (A𝑖 x, A𝑖 x0).
𝑖 𝑖

Several specific schemes have been proposed for building such de-
compositions. One convenient approach is to partition the coordinates
43 k. kandasamy et al. (2015). High Dimensional
of the input into disjoint groups and add a contribution defined on each Bayesian Optimisation and Bandits via Addi-
subset.43, 44 Figure 3.10 shows an example, where a two-dimensional ob- tive Models. icml 2015.
jective is the sum of independent axis-aligned components. We might use 44 j. r. gardner et al. (2017). Discovering and Ex-
such a model when every feature of the input is likely to be relevant but ploiting Additive Structure for Bayesian Opti-
only through interaction with a limited number of additional variables. mization. aistats 2017.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 63

 modeling with gaussian processes

An advantage of a disjoint partition is that we may reduce optimization

45 p. rolland et al. (2018). High-Dimensional
Bayesian Optimization via Additive Models of the high-dimensional objective to separate optimization of each of its
with Overlapping Groups. aistats 2018. lower-dimensional components (3.31). Several other additive schemes
46 m. mutný and a. krause (2018). Efficient have been proposed as well, including partitions with (perhaps sparsely)
High Dimensional Bayesian Optimization overlapping groups45, 46, 47 and decompositions of the general form (3.31)
with Additivity and Quadrature Fourier Fea- with arbitrary projection matrices.48, 49
tures. neurips 2018.
47 t. n. hoang et al. (2018). Decentralized High-
Dimensional Bayesian Optimization with Fac- 3.6 summary of major ideas
tor Graphs. aaai 2018.
Specifying a Gaussian process entails choosing a mean and covariance
48 e. gilboa et al. (2013). Scaling Multidimen-
sional Gaussian Processes using Projected Ad- function for the function of interest. As we saw in the previous chapter,
ditive Approximations. icml 2013. the structure of these functions has important implications regarding
49 c.-l. li et al. (2016). High Dimensional Bayes- sample path behavior, and as we will see in the next chapter, important
ian Optimization via Restricted Projection Pur- implications regarding its ability to explain a given set of data.
suit Models. aistats 2016. In practice, the design of a Gaussian process model is usually data-
driven: we establish some space of candidate models to consider, then
search this space for the models providing the best explanation of avail-
able data. In this chapter we offered some guidance for the construction
of models – or parametric spaces of models – as possible explanations of
a given system. We will continue the discussion in the next chapter by
taking up the question of assessing model quality in light of data. Below
we summarize the important ideas arising in the present discussion.
prior mean function: § 3.1, p. 46 • The mean function of a Gaussian process determines the expected value
of function values. Although an important concern, the mean function
impact on sample path behavior: figure 3.1, can only affect sample path behavior through pointwise translation,
p.46 and surrounding discussion and most interesting properties of sample paths are determined by the
covariance function instead.
impact on extrapolation: figure 3.2, p. 47 and • Nonetheless, the mean function has important implications for predic-
surrounding discussion tion, namely, in extrapolation. When making predictions in locations
poorly explained by available data – that is, locations where function
value are not strongly correlated with any observation – the prior mean
function effectively determines the posterior predictive mean.
• There are no restrictions on the mean function of a Gaussian process,
and we are free to use any sensible choice in a given scenario. In practice,
unless a better option is apparent, the mean function is usually taken to
have some relatively simple parametric form, such as a constant (3.1) or
a low-order polynomial (3.7). Such choices are both simple and unlikely
to cause grossly undesirable extrapolatory behavior.
• When the mean function includes a linear combination of basis functions,
we may exactly marginalize the coefficients under a multivariate normal
prior (3.5). The result is a marginal Gaussian process where uncertainty
in the linear terms of the mean is absorbed into the covariance function
(3.6). As an important special case, we may marginalize the value of a
constant mean (3.3) under a normal prior (3.2).
prior covariance function: § 3.2, p. 49 • The covariance function of a Gaussian process is critical to determining
the behavior of its sample paths. To be valid, a covariance function must

64
 3.6. summary of major ideas

be symmetric and positive semidefinite. The latter condition can be

difficult to guarantee for arbitrary “similarity measures,” but covariance
functions are closed under several natural operations, allowing us to
build complex covariance functions from simple building blocks.
• In particular, sums and pointwise products of covariance functions are sums and products of covariance functions:
valid covariance functions, and by extension any polynomial expression § 3.4, p. 55
of covariance functions with positive coefficients.
• Many common covariance functions are invariant to translation of their stationarity: § 3.2, p. 50
inputs, a property known as stationarity. An important result known
as bochner’s theorem provides a useful representation for the space of bochner’s theorem: § 3.2, p. 51
stationary covariance functions: their Fourier transforms are symmetric,
finite measures, and vice versa. This result has important implications for
modeling and computation, as the Fourier representation can be much
easier to work with than the covariance function itself.
• Numerous useful covariance functions are available “off-the-shelf.” The the Matérn family and squared exponential
family of Matérn covariances – and its limiting case the squared expo- covariance: § 3.3, p. 51
nential covariance – can model functions with any desired degree of
smoothness (3.11–3.14). A notable special case is the Matérn covariance
with 𝜈 = 5/2 (3.14), which has been promoted as a reasonable default.
• The spectral mixture covariance (3.15) appeals to bochner’s theorem to spectral mixture covariance: § 3.3, p. 53
provide a parametric family of covariance functions able to approximate
any stationary covariance.
• Covariance functions can be modified by arbitrary scaling of function scaling function outputs: § 3.4, p. 55
outputs (3.19) and/or arbitrary transformation of function inputs (3.21). transforming function inputs: § 3.4, p. 56
This ability allows us to create parametric families of covariance functions
with tunable behavior.
• Considering arbitrary constant scaling of function outputs gives rise to
parameters known as output scales (3.20).
• Considering arbitrary dilations of function inputs gives rise to parame-
ters known as characteristic length scales (3.22). Taking the dilation to
be anisotropic introduces a characteristic length scale for each input
dimension, a construction known as automatic relevance determination
(ard). With an ard covariance, setting a given dimension’s length scale
very high effectively “turns off” its influence on the model.
• Nonlinear warping of function inputs is also possible. This enables us to nonlinear warping: figure 3.7, p. 59 and
easily build custom nonstarionary covariance functions by combining a surrounding discussion
nonlinear warping with a stationary base covariance.
• Optimization can be especially challenging in high dimensions due to modeling functions on high-dimensional
the curse of dimensionality. However, if an objective function has intrin- domains: § 3.5, p. 61
sic low-dimensional structure, we can avoid some of the challenges by
finding a structure-preserving mapping to a lower-dimensional space
and modeling the function on the “smaller” space. This idea has repeat-
edly proven successful, and several general-purpose constructions are
available.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 65

 4
MODEL ASSESSMEN T, SELECTION, AND AVERAGING

The previous chapter offered a glimpse into the flexibility of Gaussian

processes, which can evidently model functions with a wide range of
behavior. However, a critical question remains: how we can identify
which models are most appropriate in a given situation?
The difficulty of this question is compounded by several factors. To
begin, the number of possible choices is staggering. Any function can
serve as a mean function for a Gaussian process, and we may construct prior mean function: § 3.1, p. 46
arbitrary complex covariance functions through a variety of mechanisms. prior covariance function: § 3.2, p. 49
Even if we fix the general form of the moment functions, introducing
natural parameters such as output and length scales yields an infinite output and length scales: § 3.4, p. 54
spectrum of possible models.
Further, many systems of interest act as “black boxes,” about which
we may have little prior knowledge. Before optimization, we may have
only a vague notion of which models might be reasonable for a given
objective function or how any parameters of these models should be set.
We might even be uncertain about aspects of the observation process,
such as the nature or precise scale of observation noise. Therefore, we
may find ourselves in the unfavorable position of having infinitely many
possible models to choose from and no idea how to choose!
Acquiring data, however, provides a way out of this conundrum.
After obtaining some observations of the system, we may determine
which models are the most compatible with the data and thereby establish
preferences over possible choices, a process known as model assessment. 1 The interested reader can find an overview of
Model assessment is a surprisingly complex and nuanced subject – even this rich subject in:
if we limit the scope to Bayesian methods – and no method can rightfully a. vehtari and j. ojanen (2012). A survey of
be called “the” Bayesian approach.1 In this chapter we will present one Bayesian predictive methods for model assess-
convenient framework for model assessment via Bayesian inference over ment, selection and comparison. Statistics Sur-
veys 6:142–228.
models, which are evaluated based on their ability to explain observed
data and our prior beliefs.
We will begin our presentation by carefully defining the models we models and model structures: § 4.1, p. 68
will be assessing and discussing how we may build useful spaces of mod-
els for consideration. With Gaussian processes, these spaces will most
often be built from what we will call model structures, comprising a para-
metric mean function, covariance function, and observation model; in
the context of model assessment, the parameters of these model compo-
nents are known as hyperparameters. We will then show how to perform Bayesian inference over parametric model
Bayesian inference over the hyperparameters of a model structure from spaces: § 4.2, p. 70
observations, resulting in a model posterior enabling model assessment
and other tasks. We will later extend this process to multiple model multiple model structures: § 4.5, p. 78
structures and show how we can even automatically search for better automating model structure search: § 4.6, p. 81
model structures.
Central to this approach is a fundamental measure of model fit known
as the marginal likelihood of the data or model evidence. Gaussian pro- marginal likelihood, model evidence: § 4.2,
cess models are routinely selected by maximizing this score, which can p. 71
produce excellent results when sufficient data are available to unambigu- model selection via map inference: § 4.3, p. 73
ously determine the best-fitting model. However, model construction

This material will be published by Cambridge University Press as Bayesian Optimization. 67

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 model assessment, selection, and averaging

in the context of Bayesian optimization is unusual as the expense of

gathering observations relegates us to the realm of small data. Effective
modeling with small datasets requires careful consideration of model
uncertainty: models explaining the data equally well may disagree dras-
tically in their predictions, and committing to a single model may yield
biased predictions with poorly calibrated uncertainty – and disappoint-
model averaging: § 4.4, p. 74 ing optimization performance as a result. Model averaging is one solution
that has proven effective in Bayesian optimization, where the predictions
of multiple models are combined in the interest of robustness.

4.1 models and model structures

In model assessment, we seek to evaluate a space of models according to
their ability to explain a set of observations D = (x, y). Before taking up
this problem in earnest, let us establish exactly what we mean by “model”
in this context, which is a model for the given observations, rather than
of a latent function alone as was our focus in the previous chapter.
model, 𝑝 (y | x) For this discussion we will define a model to be a prior probability
distribution over the measured values y that would result from observing
at a set of locations x: 𝑝 (y | x). In the overarching approach we have
adopted for this book, a model is specified indirectly via a prior process
on a latent function 𝑓 and an observation model linking this function to
the observed values:  
𝑝 (𝑓 ), 𝑝 (𝑦 | 𝑥, 𝜙) . (4.1)
Given explicit choices for these components, we may form the desired dis-
model induced by prior process and tribution by marginalizing the latent function values 𝝓 = 𝑓 (x) through
observation model the observation model:
∫
𝑝 (y | x) = 𝑝 (y | x, 𝝓) 𝑝 (𝝓 | x) d𝝓. (4.2)

All models we will consider below will be of this composite form (4.1), but
the assessment framework we will describe will accommodate arbitrary
models.

Spaces of candidate models

To proceed, we must establish some space of candidate models we wish
2 Although defining a space of candidate mod- to consider as possible explanations of the observed data.2 Although
els may seem natural and innocuous, this is this space can in principle be arbitrary, with Gaussian process models
actually a major point of contention between
different approaches to Bayesian model assess-
it is convenient to consider parametric collections of models defined
ment. If we subscribe to the maxim “all models by parametric forms for the observation model and the prior mean and
are wrong,” we might conclude that the true covariance functions of the latent function. We invested significant effort
model will never be contained in any space in the last chapter laying the groundwork to enable this approach: a
we define, no matter how expansive. However,
some are likely “more wrong” than others, and
running theme was the introduction of flexible parametric mean and
we can still reasonably establish preferences covariance functions that can assume a wide range of different shapes –
over the given space. perfect building blocks for expressive model spaces.
We will call a particular combination of observation model, prior
model structure mean function 𝜇, and prior covariance function 𝐾 a model structure.

68
 4.1. models and model structures

Corresponding to each model structure is a natural model space formed

by exhaustively traversing the joint parameter space: model space, M
n  o
M= 𝑝 (𝑓 | 𝜽 ), 𝑝 (𝑦 | 𝑥, 𝜙, 𝜽 ) | 𝜽 ∈ Θ , (4.3)
where


𝑝 (𝑓 | 𝜽 ) = GP 𝑓 ; 𝜇 (𝑥; 𝜽 ), 𝐾 (𝑥, 𝑥 0; 𝜽 ) .

We have indexed the space by a vector 𝜽 , the entries of which jointly vector of hyperparameters, 𝜽
specify any necessary parameters from their joint range Θ. The entries range of hyperparameter values, Θ
of 𝜽 are known as hyperparameters of the model structure, as they pa-
rameterize the prior distribution for the observations, 𝑝 (y | x, 𝜽 ) (4.2).
In many cases we may be happy with a single suitably flexible model
structure for the data, in which case we can proceed with the correspond-
ing space (4.3) as the set of candidate models. We may also consider
multiple model structures for the data by taking a discrete union of such
spaces, an idea we will return to later in this chapter. multiple model structures: § 4.5, p. 78

Example
Let us momentarily take a step back from abstraction and create an ex-
plicit model space for optimization on the interval X = [𝑎, 𝑏].3 Suppose 3 The interval can be arbitrary; our discussion
our initial beliefs are that the objective will exhibit stationary behav- will be purely qualitative.
ior with a constant trend near zero, and that our observations will be
corrupted by additive noise with unknown signal-to-noise ratio.
For the observation model, we take homoskedastic additive Gaussian observation model: additive Gaussian noise
noise, a reasonable choice when there is no obvious alternative: with unknown scale

𝑝 (𝑦 | 𝜙, 𝜎𝑛 ) = N (𝑦; 𝜙, 𝜎𝑛2 ), (4.4)

and leave the scale of the observation noise 𝜎𝑛 as a parameter. Turning prior mean function: constant mean with
to the prior process, we assume a constant mean function (3.1) with a unknown value
zero-mean normal prior on the unknown constant:

𝜇 (𝑥; 𝑐) ≡ 𝑐; 𝑝 (𝑐) = N (𝑐; 0, 𝑏 2 ),

and select the Matérn covariance function with 𝜈 = 5/2 (3.14) with un- prior covariance function: Matérn 𝜈 = 5/2
known output scale 𝜆 (3.20) and unknown length scale ℓ (3.22): with unknown output and length scales

𝐾 (𝑥, 𝑥 0; 𝜆, ℓ) = 𝜆 2 𝐾M5/2 (𝑑/ℓ).

Following our discussion in the last chapter, we may eliminate one eliminating mean parameter via
of the parameters above by marginalizing the unknown constant mean marginalization: § 3.1, p. 47
under its assumed prior,4 leaving us with the identically zero mean
function and an additive contribution to the covariance function (3.3): 4 We would ideally marginalize the other pa-
rameters as well, but it would not result in a
𝜇 (𝑥) ≡ 0; 𝐾 (𝑥, 𝑥 0; 𝜆, ℓ) = 𝑏 2 + 𝜆 2 𝐾M5/2 (𝑑/ℓ). (4.5) Gaussian process, as we will discuss shortly.

This, combined with (4.4), completes the specification of a model struc-

ture with three hyperparameters: 𝜽 = [𝜎𝑛 , 𝜆, ℓ] >
. Figure 4.1 illustrates

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 69

 model assessment, selection, and averaging

Figure 4.1: Samples from our example 𝜎𝑛

model space for a range of the
hyperparameters: 𝜎𝑛 , the ob-
servation noise scale, and ℓ,
the characteristic length scale.
The output scale 𝜆 is fixed
for each example. Each exam-
ple demonstrates a sample of
the latent function and obser-
vations resulting from mea-
surements at a fixed set of 15
locations x. Elements of the
model space can model func-
tions with short- or long-scale
correlations that are observed
with a range of fidelity from
virtually exact observation to
extreme noise. ℓ

samples from the joint prior over the objective function and the observed
values y that would result from measurements at 15 locations x (4.2) for
a range of these hyperparameters. Even this simple model space is quite
flexible, offering degrees of freedom for the variation in the objective
function and the precision of our measurements.

4.2 bayesian inference over parametric model spaces

Given a space of candidate models, we now turn to the question of
assessing the quality of these models in light of data. There are multiple
5 a. vehtari and j. ojanen (2012). A survey paths forward,5 but Bayesian inference offers one effective solution. By
of Bayesian predictive methods for model as- accepting that we can never be absolutely certain regarding which model
sessment, selection and comparison. Statistics
Surveys 6:142–228.
is the most faithful representation of a given system, we can – as with
anything unknown in the Bayesian approach – treat that “best model”
as a random variable to be inferred from data and prior beliefs.
We will limit this initial discussion to parametric model spaces built
from a single model structure (4.1), which will simplify notation and
6 As it is most likely that no model among the allow us to conflate models and their corresponding hyperparameters
candidates actually generated the data, some 𝜽 as convenient. We will consider more complex spaces comprising
authors have suggested that any choice of
multiple alternative model structures presently.
prior is dubious. If this bothers the reader, it
can help to frame the inference as being over
the model “closest to the truth” rather than
over the “true model” itself. Model prior
We first endow the model space with a prior encoding which models are
model prior, 𝑝 (𝜽 ) more plausible a priori, 𝑝 (𝜽 ).6 For convenience, it is common to design
the model hyperparameters such that the uninformative (and possibly
improper) “uniform prior”

𝑝 (𝜽 ) ∝ 1 (4.6)

70
 4.2. bayesian inference over parametric model spaces

Figure 4.2: The dataset for our model as-

sessment example, generated
using a hidden model from
the space on the facing page.

may be used, in which case the model prior may not be explicitly ac-
knowledged at all. However, it can be helpful to express at least weakly
informative prior beliefs – especially when working with small datasets
– as it can offer gentle regularization away from patently absurd choices.
This should be possible for most hyperparameters in practice. For ex-
ample, when modeling a physical system, it would be unlikely that
interaction length scales of say one nanometer and one kilometer would
be equally plausible a priori; we might capture this intuition with a wide
prior on the logarithm of the length scale.

Model posterior
Given a set of observations D = (x, y), we may appeal to Bayes’ theorem model posterior, 𝑝 (𝜽 | D)
to derive the posterior distribution over the candidate models:

𝑝 (𝜽 | D) ∝ 𝑝 (𝜽 ) 𝑝 (y | x, 𝜽 ). (4.7)

The model posterior provides support to the models most consistent

with our prior beliefs and the observed data. Consistency with the data 7 Recall that this distribution is precisely what
is encapsulated by the 𝑝 (y | x, 𝜽 ) term, the prior pdf over observations a model defines: § 4.1, p. 68.
evaluated on the actual data.7 This value is known as the model evidence model evidence, marginal likelihood,
or the marginal likelihood of the data, as it serves as a likelihood in Bayes’ 𝑝 (y | x, 𝜽 )
theorem (4.7) and, in our class of latent function models, is computed by
marginalizing the latent function values at the observed locations (4.2).

Marginal likelihood and Bayesian Occam’s razor

8 d. j. c. mackay (2003). Information Theory, In-
Model assessment becomes trivial in light of the model posterior if we ference, and Learning Algorithms. Cambridge
simply establish preferences over models according to their posterior University Press. [chapter 28]
probability. When using the uniform model prior (4.6) (perhaps implic-
itly), the model posterior is proportional to the marginal likelihood alone,
which can be then used directly for model assessment.
It is commonly argued that the model evidence encodes automatic
penalization for model complexity, a phenomenon known as Bayesian y
Occam’s razor.8 mackay outlines a simple argument for this effect by S
noting that a model 𝑝 (y | x) must integrate to unity over all possible A cartoon of the Bayesian Occam’s razor
measurements y. Thus if a “simpler” model wishes to become more effect due to mackay. Interpreting models
“complex” by putting support over a wider range of possible observations, as pdfs over measurements y, the “simple”
model in red explains datasets in S well, but
it can only do so by reducing the support for the datasets that are already
not elsewhere. The “complex” model in blue
well explained; see the illustration in the margin. explains datasets outside S better, but in S
The marginal likelihood of a given dataset can be conveniently com- worse; the probability density must be lower
puted in closed form for Gaussian process models with additive Gaussian there to explain a broader range of data.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 71

 model assessment, selection, and averaging

Figure 4.3: The posterior distribution 𝜎𝑛

over the model space from
figure 4.1 (the range of the
axes are compatible with that 2
figure) conditioned on the
dataset in figure 4.2. The out- 3
put scale is fixed (to its true
value) for the purposes of il-
lustration. Significant uncer-
tainty remains in the exact
values of the hyperparame-
ters, but the model posterior
favors models featuring either ∗
short length scales with low
noise or long length scales
with high noise. The points
marked 1–3 are referenced in 1 0 map
figure 4.4; the point marked ∗ ℓ
is the map (figure 4.5).

noise or exact observation. In this case, we have (2.18):

𝑝 (y | x, 𝜽 ) = N (y; 𝝁, 𝚺 + N),
where 𝝁 and 𝚺 are the prior mean and covariance of the latent objective
function values 𝝓 (2.3), and N is the observation noise covariance matrix
(the zero matrix for exact observation) – all of which may depend on
𝜽 . As this value can be exceptionally small and have high dynamic
marginal likelihood for Gaussian process range, the logarithm of the marginal likelihood is usually preferred for
models with additive Gaussian noise computational purposes (a.6–a.7):

log 𝑝 (y | x, 𝜽 ) =
 
− 12 (y − 𝝁) > (𝚺 + N) −1 (y − 𝝁) + log |𝚺 + N| + 𝑛 log 2𝜋 . (4.8)
interpretation of terms The first term of this expression is the sum of the squared Mahalanobis
norms (a.8) of the observations under the prior and represents a measure
of data fit. The second term serves as a complexity penalty: the volume
of any confidence ellipsoid under the prior is proportional to |𝚺 + N|,
9 The dataset was realized using a moderate
and thus this term scales according to the volume of the model’s support
length scale (30 length scales spanning the in observation space. The third term simply ensures normalization.
domain) and a small amount of additive noise,
shown below. But this is impossible to know
from inspection of the data alone, and many Return to example
alternative explanations are just as plausible
according to the model posterior! Let us return to our example scenario and model space. We invite the
reader to consider the hypothetical set of 15 observations in figure 4.2
from our example system of interest and contemplate which models
from our space of candidates in figure 4.1 might be the most compatible
with these observations.9
We illustrate the model posterior given this data in figure 4.3, where,
in the interest of visualization, we have fixed the covariance output

72
 4.3. model selection via posterior maximization

posterior mean
posterior 95% credible interval, 𝑦 posterior 95% credible interval, 𝜙

Figure 4.4: Posterior distributions given

the observed data correspond-
1 ing to the three settings
of the model hyperparam-
eters marked in figure 4.3.
Although remarkably differ-
ent in their interpretations,
each model represents is an
2 equally plausible explanation
in the model posterior. Model
1 favors near-exact observa-
tions with a short length
scale, and models 2–3 favor
large observation noise with
a range of length scales.
3

scale to its true value and set the range of the axes to be compatible
with the samples from figure 4.1. The model prior was designed to be
weakly informative regarding the expected order of magnitude of the
hyperparameters by taking independent, wide Gaussian priors on the 10 Both parameters are nonnegative, so the prior
logarithm of the observation noise and covariance length scale.10 has support on the entire parameter range.
The first observation we can make regarding the model posterior is
that it is remarkably broad, with many settings of the model hyperparam-
eters remaining plausible after observing the data. However, the model
posterior does express a preference for models with either low noise and
short length scale or high noise combined with a range of compatible
length scales. Figure 4.4 provides examples of objective function and
observation posteriors corresponding to the hyperparameters indicated
in figure 4.3. Although each is equally plausible in the posterior,11 their 11 The posterior probability density of these
explanations of the data are diverse. points is approximately 10% of the maximum.

4.3 model selection via posterior maximization

Winnowing down a space of candidate models to a single model for use
in inference and prediction is known as model selection. Model selection
becomes straightforward if we agree to rank candidates according to the
model posterior, as we may then select the maximum a posteriori (map) 12 If we only wish to find the maximum, there is
(4.7) model:12 no benefit to normalizing the posterior.

𝜽ˆ = arg max 𝑝 (𝜽 ) 𝑝 (y | x, 𝜽 ). (4.9)

𝜽

When the model prior is flat (4.6), the map model corresponds to the
maximum likelihood estimate (mle) of the model hyperparameters. Fig-

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 73

 model assessment, selection, and averaging

Figure 4.5: The predictions of the maxi-

mum a posteriori (map) model
from the example data in fig-
ure 4.2.

ure 4.5 shows the predictions made by the map model for our running
example; in this case, the map hyperparameters are in fact a reasonable
match to the parameters used to generate the example dataset.
acceleration via gradient-based optimization When the model space is defined over a continuous space of hyperpa-
rameters, computation of the map model can be significantly accelerated
via gradient-based optimization. Here it is advisable to work in the log
domain, where the objective becomes the unnormalized log posterior:
log 𝑝 (𝜽 ) + log 𝑝 (y | x, 𝜽 ). (4.10)
The log marginal likelihood is given in (4.8), noting that 𝝁, 𝚺, and N
are all implicitly functions of the hyperparameters 𝜽 . This objective
gradient of log marginal likelihood with (4.10) is differentiable with respect to 𝜽 assuming the Gaussian process
respect to 𝜽 : § c.1, p. 307 prior moments, the noise covariance, and the model prior are as well, in
which case we may appeal to off-the-shelf gradient methods for solving
(4.9). However, a word of warning is in order: the model posterior is
not guaranteed to be concave and may have multiple local maxima, so
multistart optimization is prudent.

4.4 model averaging

Reliance on a single model is questionable when the model posterior is
not well determined by the data. For example, in our running example,
a diverse range of models are consistent with the data (figures 4.3–4.4).
Committing to a single model in this case may systematically bias our
predictions and underestimate predictive uncertainty – note how the
diversity in predictions from figure 4.4 is lost in the map model (4.5).
model-marginal objective posterior, 𝑝 ( 𝑓 | D) An alternative is to marginalize the model with respect to the model
model-marginal predictive distribution, posterior, a process known as model averaging:
𝑝 (𝑦 | 𝑥, D) ∫
𝑝 (𝑓 | D) = 𝑝 (𝑓 | D, 𝜽 ) 𝑝 (𝜽 | D) d𝜽 ; (4.11)
∬
𝑝 (𝑦 | 𝑥, D) = 𝑝 (𝑦 | 𝑥, 𝜙, 𝜽 ) 𝑝 (𝜙 | 𝑥, D, 𝜽 ) 𝑝 (𝜽 | D) d𝜙 d𝜽, (4.12)

where we have marginalized the hyperparameters of both the objective

and observation models. Model averaging is more consistent with the
ideal Bayesian convention of marginalizing nuisance parameters when
13 Although it may be unusual to consider the possible13 and promises robustness to model misspecification, at least
choice of model a “nuisance!” over the chosen model space.
Unfortunately, neither of these model-marginal distributions (4.11–
4.12) can be computed exactly for Gaussian process models except in

74
 4.4. model averaging

posterior mean posterior 95% credible interval, 𝑦 posterior 95% credible interval, 𝜙

Figure 4.6: A Monte Carlo estimate to the model-marginal predictive distribution (4.11) for our example sceneario using
100 samples drawn from the model posterior in figure 4.3 (4.14–4.15); see illustration in margin. Samples
from the objective function posterior display a variety of behavior due to being associated with different
hyperparameters.

some special cases,14 so we must resort to approximation if we wish 14 A notable example is marginalizing the coeffi-
to pursue this approach. In fact, maximum a posteriori estimation can cients of a linear prior mean against a Gaus-
sian prior: § 3.1, p. 47.
be interpreted as one rather crude approximation scheme where the
model posterior is replaced by a Dirac delta distribution at the map
hyperparameters:
𝑝 (𝜽 | D) ≈ 𝛿 (𝜽 − 𝜽ˆ ).
This can be defensible when the dataset is large compared to the number
of hyperparameters, in which case the model posterior is often unimodal
with little residual uncertainty. However, large datasets are the exception
rather than the rule in Bayesian optimization, and more sophisticated
approximations can pay off when model uncertainty is significant.

Monte Carlo approximation

Monte Carlo approximation is one straightforward path forward. Draw-
ing a set of hyperparameter samples from the model posterior,

{𝜽𝑖 }𝑠𝑖=1 ∼ 𝑝 (𝜽 | D), (4.13)

yields the following simple Monte Carlo estimates:

1 ∑︁

𝑠
𝑝 (𝑓 | D) ≈ GP 𝑓 ; 𝜇D (𝜽𝑖 ), 𝐾D (𝜽𝑖 ) ; (4.14) 𝜎𝑛
𝑠 𝑖=1
𝑠 ∫
1 ∑︁
𝑝 (𝑦 | 𝑥, D) ≈ 𝑝 (𝑦 | 𝑥, 𝜙, 𝜽𝑖 ) 𝑝 (𝜙 | 𝑥, D, 𝜽𝑖 ) d𝜙 . (4.15)
𝑠 𝑖=1

The objective function posterior is approximated by a mixture of Gaus-

sian processes corresponding to the sampled hyperparameters, and the
posterior predictive distribution for observations is then derived by inte- ℓ
grating a Gaussian mixture (2.36) against the observation model.
Any Markov chain Monte Carlo procedure could be used to generate The 100 hyperparameter samples used to
the hyperparameter samples (4.13); a variation on Hamiltonian Monte produce figure 4.6.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 75

 model assessment, selection, and averaging

posterior mean posterior 95% credible interval, 𝑦 posterior 95% credible interval, 𝜙

Figure 4.7: An approximation to the model-marginal posterior (4.11) using the central composite design approach proposed
by rue et al. A total of nine hyperparameter samples are used for the approximation, illustrated in the margin
below.

15 m. d. hoffman and a. gelman (2014). The Carlo (hmc) such as the no u-turn sampler (nuts) would be a reasonable
No-U-turn Sampler: Adaptively Setting Path choice when the gradient of the log posterior (4.10) is available, as it can
Lengths in Hamiltonian Monte Carlo. Journal
of Machine Learning Research 15(4):1593–1623.
exploit this information to accelerate mixing.15
Figure 4.6 demonstrates a Monte Carlo approximation to the model-
marginal posterior (4.11–4.12) for our running example. Comparing with
the map approximation in figure 4.5, the predictive uncertainty of both
objective function values and observations has increased considerably
due to accounting for model uncertainty in the predictive distributions.

Deterministic approximation schemes

The downside of Monte Carlo approximation is relatively inefficient use
of the hyperparameter samples – the price of random sampling rather
than careful design. This inefficiency in turn leads to an increased com-
putational burden for inference and prediction from having to derive
a gp posterior for each sample. Several more efficient (but less accu-
rate) alternative approximations for hyperparameter marginalization
have also been proposed. A common simplifying tactic taken by these
cheaper procedures is to approximate the hyperparameter posterior with
Laplace approximation: § b.1, p. 301 a multivariate normal via a Laplace approximation:
ˆ C),
𝑝 (𝜽 | D) ≈ N (𝜽 ; 𝜽, (4.16)
where 𝜽ˆ is the map (4.9). Integrating this approximation into (4.11) gives
∫


𝑝 (𝑓 | D) ≈ GP 𝑓 ; 𝜇D (𝜽 ), 𝐾D (𝜽 ) N (𝜽 ; 𝜽,ˆ C) d𝜽 . (4.17)

16 h. rue et al. (2009). Approximate Bayesian Unfortunately this integral remains intractable due to the nonlinear de-
inference for latent Gaussian models by us- pendence of the posterior moments on the hyperparameters, but reducing
ing integrated nested Laplace approximations.
Journal of the Royal Statistical Society Series B to this common form allows us to derive deterministic approximations
(Methodological) 71(2):319–392. against a single assumed posterior.
17 g. e. p. box and k. b. wilson (1951). On the Ex- rue et al. introduced several approximation schemes representing
perimental Attainment of Optimum Condi- different tradeoffs between efficiency and fidelity.16 Notable among these
tions. Journal of the Royal Statistical Society is a simple, sample-efficient procedure grounded in classical experi-
Series B (Methodological) 13(1):1–45.
mental design. Here a central composite design17 in hyperparameter

76
 4.4. model averaging

posterior mean posterior 95% credible interval, 𝑦 posterior 95% credible interval, 𝜙

Figure 4.8: The approximation to the model-marginal posterior (4.11) for our running example using the approach proposed
by osborne et al.

space is transformed to agree with the moments of (4.16), then used as

𝜎𝑛
nodes in a numerical quadrature approximation to (4.17). The resulting
approximation again takes the form of a (now weighted) mixture of
Gaussian processes (4.14): the map model augmented by a small num-
ber of additional models designed to reflect the important variation in
the hyperparameter posterior. The number of hyperparamater samples
required by this scheme grows relatively slowly with the dimension of
the hyperparameter space: less than 100 for |𝜽 | ≤ 8 and less than 1000
for |𝜽 | ≤ 21.18 The nine samples required for our running example are ℓ
shown in the marginal figure. Figure 4.7 shows the resulting approximate
A Laplace approximation to the model pos-
posterior; comparing with the gold-standard Monte Carlo approximation
terior (performed in the log domain) and
from figure 4.6, the agreement is excellent. hyperparameter settings corresponding to the
An even more lightweight approximation was proposed by osborne central composite design proposed by rue
et al., which despite its crudeness is arguably still preferable to map et al. The samples do a good job covering the
support of the true posterior.
estimation and can be used as a drop-in replacement.19 This approach
again relies on a Laplace approximation to the hyperparameter posterior
(4.16–4.17). The key observation is that under the admittedly strong 18 s. m. sanchez and p. j. sanchez (2005). Very
assumption that the posterior mean were in fact linear in 𝜽 and the Large Fractional Factorial and Central Com-
posite Designs. acm Transactions on Modeling
posterior covariance independent of 𝜽, we could resolve (4.17) in closed and Computer Simulation 15(4):362–377.
form. We proceed by taking the best linear approximation to the posterior
19 m. a. osborne et al. (2012). Active Learning of
mean around the map:20 Model Evidence Using Bayesian Quadrature.
neurips 2012.
𝜕𝜇D (𝑥; 𝜽 ) ˆ
𝜇D (𝑥; 𝜽 ) ≈ 𝜇D (𝑥; 𝜽ˆ ) + g(𝑥) >(𝜽 − 𝜽ˆ ); g(𝑥) = (𝜽 ), 20 This is analogous to the linearization step in
𝜕𝜽 the extended Kalman filter, whereas the cen-
tral composite design approach is closer to the
and assuming the map posterior covariance is universal: 𝐾D (𝜽 ) ≈ 𝐾D (𝜽ˆ ). unscented Kalman filter in pushing samples
The result is a single Gaussian process approximation to the posterior: through the nonlinear transformation.

𝑝 (𝑓 | D) ≈ GP (𝑓 ; 𝜇ˆD , 𝐾ˆD ), (4.18)

where

𝜇ˆD (𝑥) = 𝜇D (𝑥; 𝜽ˆ ); 𝐾ˆD (𝑥, 𝑥 0) = 𝐾D (𝑥, 𝑥 0; 𝜽ˆ ) + g(𝑥) >Cg(𝑥 0).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 77

 model assessment, selection, and averaging

This is the map model with covariance inflated by a term determined by

the dependence of the posterior mean on the hyperparameters, g, and
the uncertainty in the hyperparameters, C.21
uncertainty in additive noise scale 𝜎𝑛 osborne et al. did not address how to account for uncertainty in
observation model parameters when approximating 𝑝 (𝑦 | 𝑥, D), but we
21 This term vanishes if the hyperparameters are can derive a natural approach for independent additive Gaussian noise
completely determined by the data, in which with unknown scale 𝜎𝑛 . Given 𝑥, let 𝑝 (𝜙 | 𝑥, D) ≈ N (𝜙; 𝜇, 𝜎 2 ) as in
case the approximation regresses gracefully
to the map estimate. (4.18). We must approximate22
∫
22 In general we have
𝑝 (𝑦 | 𝑥, D) ≈ N (𝑦; 𝜇, 𝜎 2 + 𝜎𝑛2 ) 𝑝 (𝜎𝑛 | 𝑥, D) d𝜎𝑛 .
𝑝 (𝑦 | 𝑥, D) =
∬
𝑝 (𝑦 | 𝑥, 𝜙, 𝜎𝑛 ) 𝑝 (𝜙, 𝜎𝑛 | 𝑥, D) d𝜙 d𝜎𝑛 ,
A moment-matched approximation 𝑝 (𝑦 | 𝑥, D) ≈ N (𝑦; 𝑚, 𝑠 2 ) is possible
by appealing to the law of total variance:
and we have resolved the integral on 𝜙 using
the single-gp approximation. 𝑚 = 𝔼[𝑦 | 𝑥, D] ≈ 𝜇; 𝑠 2 = var[𝑦 | 𝑥, D] ≈ 𝜎 2 + 𝔼[𝜎𝑛2 | 𝑥, D].
If the noise scale is parameterized by its logarithm, then the Laplace
approximation (4.16) in particular yields
𝑝 (log 𝜎𝑛 | 𝑥, D) ≈ N (log 𝜎𝑛 ; log 𝜎ˆ𝑛 , 𝑠 2 ); 𝔼[𝜎𝑛2 | 𝑥, D] ≈ 𝜎ˆ𝑛2 exp(2𝑠 2 ).
Thus we predict with the map estimate 𝜎ˆ𝑛 inflated by a factor commen-
surate with the residual uncertainty in the noise contribution.
Figure 4.8 shows the resulting approximation for our running exam-
ple. Although not perfect, the predictive uncertainty in the observations
is more faithful than the map model from figure 4.5, which severely
underestimates the scale of observation noise in the posterior.

4.5 multiple model structures

We have now covered model inference, selection, and averaging with
a single parametric model space (4.1). With a bit of extra bookkeeping,
we may extend this framework to handle multiple model structures
comprising different combinations of parametric prior moments and
observation models.
To begin, we may build a space of candidate models by taking a
discrete union of parametric spaces as in (4.1), with one built from each
model structure index, M desired model structure: {M𝑖 }. It is natural to index this space by (𝜽, M),
where 𝜽 is understood to be a vector of hyperparameters associated with
the specified model structure; the size and interpretation of this vector
may differ across structures. All that remains is to derive our previous
results while managing this compound structure–hyperparameter index.
We may define a model prior over this compound space by com-
bining a prior over the chosen model structures with priors over the
model structure prior, Pr(M) hyperparameters of each:
𝑝 (𝜽, M) = Pr(M) 𝑝 (𝜽 | M). (4.19)
Given data, the model posterior has a similar form as before (4.7):
𝑝 (𝜽, M | D) = Pr(M | D) 𝑝 (𝜽 | D, M). (4.20)

78
 4.5. multiple model structures

Figure 4.9: The objective and dataset for

our multiple-model example.

The structure-conditional hyperparameter posterior 𝑝 (𝜽 | D, M) is as

in (4.7) and may be reasoned about following our previous discussion.
The model structure posterior is then given by model structure posterior, Pr(M | D)

Pr(M | D) ∝ Pr(M) 𝑝 (y | x, M); (4.21)

∫
𝑝 (y | x, M) = 𝑝 (y | x, 𝜽, M) 𝑝 (𝜽 | M) d𝜽 . (4.22)

The expression in (4.22) is the normalizing constant of the structure-

conditional hyperparameter posterior (4.7), which we could ignore when
there was only a single model structure. This integral is in general
intractable, but several approximations are feasible. One effective choice
is the Laplace approximation (4.16), which provides an approximation Laplace approximation: § b.1, p. 301
to the integral as a side effect (b.2). The classical Bayesian information
criterion (bic) may be seen as an approximation to this approximation.23 23 s. konishi and g. kitagawa (2008). Informa-
Model selection may now be pursued by maximizing the model tion Criteria and Statistical Modeling. Springer–
Verlag. [chapter 9]
posterior over the model space as before, although we may no longer
appeal to gradient methods as the model space is not continuous with model selection
multiple model structures. A simple approach would be to find the map
hyperparameters for each of the model structures separately, then use
these map points to approximate (4.22) for each structure via the Laplace
approximation or bic. This would be sufficient to estimate (4.20–4.21)
and maximize over the map models.
Turning to model averaging, the model-marginal posterior to the model averaging
objective function is:
∑︁
𝑝 (𝑓 | D) = Pr(M𝑖 | D) 𝑝 (𝑓 | D, M𝑖 ). (4.23)
𝑖

The structure-conditional, hyperparameter-marginal distribution on

each space 𝑝 (𝑓 | D, M) is as before (4.11) and may be approximated
following our previous discussion. These are now combined in a mixture
distribution weighted by the model structure posterior (4.21).

Multiple structure example

We now present an example of model inference, selection, and averaging
over multiple model structures using the dataset in figure 4.9.24 The data 24 The data are used as a demo in the code re-
were sampled from a Gaussian process with linear prior mean (a linear leased with:
trend with positive slope is evident) and Matérn 𝜈 = 3/2 prior covariance c. e. rasmussen and c. k. i. williams (2006).
(3.13), with a small amount of additive Gaussian noise. We also show a Gaussian Processes for Machine Learning. mit
sample from the objective function posterior corresponding to the true Press.
model generating the data for reference.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 79

 model assessment, selection, and averaging

m5 lin lin × lin

Figure 4.10: Sample paths from our ex-

ample model structures.

m5 + lin m5 × lin

initial model structure: p. 69 We build a model space comprising several model structures by aug-
menting our previous space with structures incorporating additional
covariance functions. The treatment of the prior mean (unknown con-
stant marginalized against a Gaussian prior) and observation model
(additive Gaussian noise with unknown scale) will remain the same
for all. The model structures reflect a variety of hypotheses positing
potential linear or quadratic behavior:
m5: the Matérn 𝜈 = 5/2 covariance (3.14) from our previous example;
lin: the linear covariance (3.16), where the the prior on the slope is vague
and centered at zero and the prior on the intercept agrees with the m5
model;
lin × lin: the product of two linear covariances designed as above, modeling a
latent quadratic function with unknown coefficients;
m5 + lin: the sum of a Matérn 𝜈 = 5/2 and linear covariance designed as in the
corresponding individual model structures; and
m5 × lin: the product of a Matérn 𝜈 = 5/2 and linear covariance designed as in the
corresponding individual model structures.
Objective function samples from models in each of these structures are
shown in figure 4.10. Among these, the model structure closest to the
truth is arguably m5 + lin.
approximation to model structure posterior Following the above discussion, we find the map hyperparameters
for each of these model structures separately and use a Laplace approx-
imation (4.16) to approximate the hyperparameter posterior on each
space, along with the normalizing constant (4.22). Normalizing over the
structures provides an approximate model structure posterior:

Pr(m5 | D) ≈ 10.8%;
Pr(m5 + lin | D) ≈ 71.8%;
Pr(m5 × lin | D) ≈ 17.0%,

with the remaining model structures (lin and lin × lin) sharing the re-

80
 4.6. automating model structure search

posterior mean objective

posterior 95% credible interval, 𝜙

Figure 4.11: An approximation to the model-marginal posterior (4.23) for our multiple-model example. The posterior on
each model structure is approximated separately as a mixture of Gaussian processes following rue et al. (see
figure 4.7); these are then combined by weighting by an approximation of the model structure posterior (4.21).
We show the result with three superimposed, transparent credible intervals, which are shaded with respect to
their weight in contributing to the final approximation.

maining 0.4%. The m5 + lin model structure is the clear winner, and there
is strong evidence that the purely polynomial models are insufficient for
explaining the data alone.
Figure 4.11 illustrates an approximation to the model-marginal pos- approximation to marginal predictive
terior (4.23), approximated by applying rue et al.’s central composite distribution
design approach to each of the model structures, then combining these
into a Gaussian process mixture by weighting by the approximate model
structure posterior. The highly asymmetric credible intervals reflect
the diversity in explanations for the data offered by the chosen model
structures, and the combined model makes reasonable predictions of our
example objective function sampled from the true model.
For this example, averaging over the model structure has important averaging over a space of Gaussian processes
implications regarding the behavior of the resulting optimization policy. in policy computation: § 8.10, p. 192
Figure 4.12 illustrates a common acquisition function25 built from the off- 25 specifically, expected improvement: § 7.3 p. 127
the-shelf m5 model, as well as from the structure-marginal model. The
former chooses to exploit near what it believes is a local optimum, but
the latter has a strong belief in an underlying linear trend and chooses
to explore the right-hand side of the domain instead. For our example
objective function sample, this would in fact reveal the global optimum
with the next observation.

4.6 automating model structure search

We now have a comprehensive framework for reasoning about model
uncertainty, including methods for model assessment, selection, and
averaging across one or multiple model structures. However, it is still not
clear how we should determine which model structures to consider for a
given system. This is critical as our model inference procedure requires

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 81

 model assessment, selection, and averaging

acquisition function next observation location

Figure 4.12: Optimization policies built from the map m5 model (left) and the structure-marginal posterior (right). The m5
model chooses to exploit near the local optimum, but the structure-marginal model is aware of the underlying
linear trend and chooses to explore the right-hand side as a result.

the space of candidate models to be predefined; equivalently, the model

prior (4.19) is implicitly set to zero for every model outside this space.
Ideally, we would simply enumerate every possible model structure and
average over all of them, but even a naïve approximation of this ideal
would entail overwhelming computational effort.
However, the set of model structures we consider for a given dataset
can be adaptively tailored as we gather data. One powerful idea is to
appeal to metaheuristics such as local search: by establishing a suitable
space of candidate model structures, we can dynamically explore this
space for the best explanations of available data.

Spaces of candidate model structures

26 d. duvenaud et al. (2013). Structure Discovery To enable this approach, we must first establish a sufficiently rich space
in Nonparametric Regression through Com- of candidate model structures. duvenaud et al. proposed one convenient
positional Kernel Search. icml 2013.
mechanism for defining such a space via a simple productive grammar.26
addition and multiplication of covariance The idea is to appeal to the closure of covariance functions under ad-
functions: § 3.4, p. 60 dition and pointwise multiplication to systematically build up families
of increasingly complex models from simple components. We begin by
base kernels, B choosing a set of so-called base kernels, B, modeling relatively simple
behavior, then extend this set to an infinite family of compositions via
the following context-free grammar:

𝐾 →𝐵
𝐾 →𝐾 +𝐾
𝐾 → 𝐾𝐾
𝐾 → (𝐾).

The symbol 𝐵 in the first rule represents any desired base kernel. The five
model structures considered in our multiple-structure example above in

82
 4.6. automating model structure search

fact represent five members of the language generated by this grammar

with the base kernels B = {𝐾M5/2, 𝐾lin } or simply B = {m5, lin}. The
grammar however also generates arbitrarily more complicated expres-
sions such as


m5 + (m5 + lin)m5 (m5 + m5) + lin. (4.24)
We are free to design the base kernels to capture any potential atomic
behavior in the objective function. For example, if the domain is high
dimensional and we suspect that the objective may depend only on Samples from objective function models
sparse interactions of mostly independent variables, we might design the incorporating the example covariance
base kernels to model variations in single variables at a time, then rely structure (4.24).
on the grammar to generate an array of possible interaction structures.
Other spaces of model structures have also been proposed for auto- additive decompositions, § 3.5, p. 61
mated structure search. With an eye toward high-dimensional domains,
gardner et al. for example considered spaces of additive model struc-
tures indexed by every possible partition of the input variables.27 This 27 j. r. gardner et al. (2017). Discovering and Ex-
is an expressive class of model structures, but the number of partitions ploiting Additive Structure for Bayesian Opti-
mization. aistats 2017.
grows so rapidly that exhaustive search is not feasible.

Searching over model structures

Once a space of candidate model structures has been established, we
may develop a search procedure seeking the most promising structures
to explain a given dataset. Several approaches have been proposed for
this search with a range of complexity, all of which frame the problem in
terms of optimizing some figure of merit over the space. Although any
score could be used in this context, a natural choice is an approximation
to the (unnormalized) model structure posterior (4.21) such as the Laplace
approximation or the Bayesian information criterion, and every method
described below uses one of these two scores.
duvenaud et al. suggested traversing their kernel grammar via
greedy search – here, we first evaluate the base kernels, then subject the
productive rules to the best among them to generate similar structures to
search next.26 We continue in this manner as desired, alternating between
evaluating the newly proposed structures, then using the grammar to
expand around the best-seen structure to generate new proposals. This
simple procedure is easy to implement and offers a strong baseline.
malkomes et al. refined this approach by replacing greedy search
with Bayesian optimization over the space of model structures.28 As in 28 g. malkomes et al. (2016). Bayesian optimiza-
the duvenaud et al. procedure, the authors pose the problem in terms of tion for automated model selection. neurips
maximizing a score over model structures: a Laplace approximation of
2016.

the (log) unnormalized structure posterior (4.21). This objective function

was then modeled using a Gaussian process, which informed a sequen-
tial Bayesian optimization procedure seeking to effectively manage the
exploration–exploitation tradeoff in the space of candidate structures.
The Gaussian process in model space requires a covariance function over
model structures, and the authors proposed an exotic “kernel kernel”
evaluating the similarity of proposed structures in terms of the overlap

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 83

 model assessment, selection, and averaging

between their hyperparameter-marginal priors for the given dataset. The

resulting optimization procedure was found to rapidly locate promising
models across a range of regression tasks.

End-to-end automation
Follow-on work demonstrated a completely automated Bayesian opti-
mization system built on this structure search procedure avoiding any
29 g. malkomes and r. garnett (2018). Automat- manual modeling at all.29 The key idea was to dynamically maintain a
ing Bayesian optimization with Bayesian opti- set of plausible model structures throughout optimization. Predictions
mization. neurips 2018.
are made via model averaging over this set, offering robustness to model
misspecification when computing the outer optimization policy. Every
time a new observation is obtained, the set of model structures is then
updated via a continual Bayesian optimization in model space given the
new data. This interleaving of Bayesian optimization in data space and
model space offered promising performance.
Finally, gardner et al. offered an alternative to optimization over
model structures by constructing a Markov chain Monte Carlo routine
30 j. r. gardner et al. (2017). Discovering and Ex- to sample model structures from their posterior (4.21).30 The proposed
ploiting Additive Structure for Bayesian Opti- sampler was a realization of the Metropolis–Hastings algorithm with a
mization. aistats 2017.
custom proposal distribution making minor modifications to the incum-
bent structure. In the case of the additive decompositions considered
in that work, this step consisted of applying random atomic operations
such as merging or splitting components of the existing decomposition.
Despite the absolutely enormous number of possible additive decomposi-
tions, this mcmc routine was able to quickly locate promising structures,
and averaging over the sampled structures for prediction resulted in
superior optimization performance as well.

4.7 summary of major ideas

We have presented a convenient framework for model assessment, se-
lection, and averaging grounded in Bayesian inference; this is the pre-
dominant approach with Gaussian process models. In the context of
31 j. snoek et al. (2012). Practical Bayesian Op- Bayesian optimization, perhaps the most important development was
timization of Machine Learning Algorithms. the notion of model averaging, which has proven beneficial to empirical
neurips 2012.
performance31 and has become standard practice.
• Model assessment entails deriving preferences over a space of candidate
models of a given system in light of available data.
models and model structures: § 4.1, p. 68 • In its purest form, a model in this context is a prior distribution over
observed values y arising from observations at a given set of locations x,
𝑝 (y | x). A convenient mechanism for specifying a model is via a prior
process for a latent function, 𝑝 (𝑓 ), and an observation model conditioned
on this function, 𝑝 (𝑦 | 𝑥, 𝜙) (4.1–4.2).
• With Gaussian process models, it is convenient to work with combina-
tions of parametric forms for the prior mean function, prior covariance

84
 4.7. summary of major ideas

function, and observation model, a construct we call a model structure. A

model structure defines a space of corresponding models by traversing
its parameter space (4.3), allowing us to build expressive model spaces.
• Once we delineate a space of candidate models, model assessment be- Bayesian inference over (parametric) model
comes straightforward if we make the – perhaps dubious but nonetheless spaces: § 4.2, p. 70
practical – assumption that the mechanism generating our data is con-
tained within this space. This allows us to treat that true model as a
random variable and proceed via Bayesian inference over the chosen
model space.
• This inference proceeds as usual. We first define a model prior capturing
any initial beliefs over the model space. Then, given a set of observations
D = (x, y), the model posterior is proportional to the model prior and a
measure of model fit known as the marginal likelihood or model evidence,
the probability (density) of the observed data under the model.
• In addition to quantifying fit, the model evidence encodes an automatic Bayesian Occam’s razor, § 4.2, p. 71
penalty for model complexity, an effect known as Bayesian Occam’s razor.
• Model evidence can be computed in closed form for Gaussian process
models with additive Gaussian observation noise (4.8).
• Model inference is especially convenient when the model space is a multiple model structures: § 4.5, p. 78
single model structure, but can be extended to spaces built from multiple
model structures with a bit of extra bookkeeping.
• The model posterior provides a simple means of model assessment by model selection via map inference: § 4.3, p. 73
establishing preferences according to posterior probability. If we must
commit to a single model to explain the data – a task known as model
selection – we then select the maximum a posteriori (map) model.
• Model selection may not be prudent when the model posterior is very model averaging: § 4.4, p. 74
flat, which is common when observations are scarce. In this case many
models may be compatible with the data but incompatible in their predic-
tions, which should be accounted for in the interest of robustness. Model
averaging is a natural solution, where we marginalize the unknown
model when making predictions according to the model posterior.
• Model averaging cannot in general be performed in closed form for approximations to model-marginal posterior:
Gaussian process models; however, we may proceed via mcmc sampling figures 4.6–4.8 and surrounding text
(4.14–4.15) or by appealing to more lightweight approximation schemes.
• Appealing to metaheuristics allows us to automatically search a space of automating model structure search: § 4.6, p. 81
candidate model structures to explain a given dataset. Once sufficiently
mature, such schemes may some day enable fully automated Bayesian
optimization pipelines that sidestep explicit modeling altogether.
The next chapter marks a major departure from our discussion thus
far, which has focused on modeling and making predictions from data. We
will now shift our attention from inference to decision making, with the
goal of building effective optimization policies informed by the models
we have now fully developed. This endeavor will consume the bulk of
the remainder of the book.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 85

model assessment, selection, and averaging

The first step will be to develop a framework for optimal decision

making under uncertainty. Our work to this point will serve an essential
component of this framework, as every such decision will be made with
reference to a posterior belief about what might happen as a result. In
the context of optimization, this belief will take the form of a posterior
predictive distribution for proposed observations given data, and our
investment in building faithful models will pay off in spades.

86
 5
DECISION THEORY FOR OPTIMIZATION

Optimization entails a series of decisions. Most obviously, we must re-

peatedly decide where to make each observation guided by the available
data. Some settings also demand we decide when to terminate opti-
mization, weighing the potential benefit from continuing optimization
against any costs that may be incurred. It is not obvious how we should
make these decisions, especially in the face of incomplete and constantly
evolving knowledge about the objective function that is only refined via
the outcomes of our own actions.
In the previous four chapters, we established Bayesian inference as a
framework for reasoning about uncertainty that offers partial guidance.
The primary obstacle to decision making during optimization is uncer-
tainty about the objective function, and, by extension, the outcomes of
proposed observations. Bayesian inference allows us to reason about an
unknown objective function with a probability distribution over plausible
functions that we may seamlessly update as we gather new informa-
tion. This belief over the objective function in turn enables prediction of
proposed observations via the posterior predictive distribution.
How can we use these beliefs to guide our decisions? Bayesian in-
ference offers no direct answer, but in this chapter we will bridge this
gap. We will develop Bayesian decision theory as a principled means Bayesian decision theory
of decision making under uncertainty and apply this approach in the
context of optimization, demonstrating how to use a probabilistic belief
about an objective function to inform intelligent optimization policies.
Recall our model of sequential optimization outlined in algorithm formalization of optimization, § 1.1, p. 2
1.1, repeated for convenience on the following page. We begin with an
arbitrary set of data, which we build upon through a sequence of obser-
vations of our own design. The core of the procedure is an optimization optimization policy
policy, which examines any already gathered data and makes the funda-
mental decision of where to make the next observation. With a policy
in hand, optimization proceeds by repeating a straightforward pattern:
the policy selects the next observation location, then we acquire the
requested measurement and update our data accordingly. We repeat this
process until satisfied, at which point we return the collected data.
Barring the question of termination, the behavior of this procedure is
entirely determined by the policy, and constructing optimization policies
will be our primary concern in this and the following chapters. We will
begin with sheer audacity: we will derive the optimal policy – in terms
of maximizing the expected quality of the returned data – in a generic optimal optimization policies: § 5.2, p. 91
setting. The reader may wonder why this book is so long if the optimal
policy is apparently so simple. As it turns out, this theoretically opti- running time and approximation: § 5.3, p. 99
mal procedure is usually impossible to compute and rarely of practical chapter 7: common Bayesian optimization
value. However, our careful derivation will shed light on how we might policies, p. 123
derive effective approximations. This is a common theme in Bayesian
chapter 8: computation of Bayesian
optimization and will be our focus in chapters 7 and 8. optimization policies for Gaussian processes,
The question of when to terminate optimization also represents p. 157
a decision that can be of critical importance in some applications. A

This material will be published by Cambridge University Press as Bayesian Optimization. 87

This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 decision theory for optimization

Algorithm 1.1: Sequential optimization. input: initial dataset D I can be empty

repeat
𝑥 ← policy(D) I select the next observation location
𝑦 ← observe(𝑥)
 I observe at the chosen location
D ← D ∪ (𝑥, 𝑦) I update dataset
until termination condition reached I e.g., budget exhausted
return D

procedure for inspecting an observed dataset and deciding whether to

stopping rule stop or continue optimization is called a stopping rule. In optimization,
the stopping rule is often fixed and known before we begin, in which case
1 A predominant example is a preallocated bud- we do not need to worry over its design.1 However, in some scenarios, we
get on the number of allowed observations, may wish instead to consider our evolving understanding of the objective
in which case we are compelled to stop after
exhausting the budget regardless of progress. function and the expected cost of further observations to dynamically
decide when to stop, requiring more subtle adaptive stopping rules. We
optimal stopping rules: § 5.4, p. 103 will also address termination decisions in this chapter and will again
begin by deriving the optimal – but intractable – stopping procedure,
practical stopping rules: § 9.3, p. 210 which will inspire efficient and effective approximations.
Practical optimization routines will return datasets that reflect signif-
icant progress on our global optimization problem (1.1) in some way. For
example, we may wish to return datasets containing near-optimal values
of the objective function. Alternatively, we may be satisfied returning
datasets that indirectly reveal likely locations of the global optimum
or achieve some other related goal. We will formalize this notion of a
utility functions for optimization: chapter 6, returned dataset’s utility shortly and use it to guide optimization. First
p. 109 we pause to introduce a useful and pervasive technique for implicitly
defining an optimization policy by maximizing a score function over the
domain.

Defining optimization policies via acquisition functions

A convenient mechanism for defining an optimization policy is by first
specifying an intermediate so-called acquisition function (also called an
acquisition function, infill function, figure of infill function or figure of merit) that provides a score to each potential
merit observation location commensurate with its propensity for aiding the
optimization task. We may then define a policy by observing at a point
judged most promising by the acquisition function. Nearly all Bayesian
optimization policies are defined in this manner, and this relationship
is so intimate that the phrase “acquisition function” is often used in-
terchangeably with “policy” in the literature and conversation, with
maximization of the acquisition function understood.
Specifically, an acquisition function 𝛼 : X → ℝ assigns a score to
each point in the domain reflecting our preferences over locations for the
next observation. Of course, these preferences will presumably depend
on the data we have already observed. To make this dependence explicit,
acquisition function, 𝛼 (𝑥; D) we adopt the notation 𝛼 (𝑥; D) for a general acquisition function, where
available data serve as parameters shaping our preferences. In the Bayes-

88
 5.1. introduction to bayesian decision theory

ian approach, acquisition functions are invariably defined by deriving

the posterior belief of the objective function given the data, 𝑝 (𝑓 | D),
then defining preferences with respect to this belief.
An acquisition function 𝛼 encodes preferences over potential obser- encoding preferences with an acquisition
vation locations by inducing a total order over the domain: given data function
D, observing at a point 𝑥 is preferred to at another point 𝑥 0 whenever
𝛼 (𝑥; D) > 𝛼 (𝑥 0; D). Thus a rational action in light of these preferences
is (any) one maximizing the acquisition function:2 2 Ties may be broken arbitrarily.

𝑥 ∈ arg max 𝛼 (𝑥 0; D). (5.1)

𝑥 0 ∈X

Solving (5.1) maps a set of observed data D to a point 𝑥 ∈ X to observe

next, exactly the role of an optimization policy.
At first this idea may sound absurd: we have proposed solving a global the paradox of Bayesian optimization: global
optimization problem (1.1) by repeatedly solving global optimization prob- optimization via. . . global optimization?
lems (5.1)! To resolve this apparent paradox, we note that acquisition
functions in common use have properties rendering their optimization
considerably more tractable than the problem we ultimately wish to
solve. Typical acquisition functions are both cheap to evaluate and ana-
lytically differentiable, allowing the use of off-the-shelf optimizers when
computing the policy (5.1). The objective function, on the other hand, is
assumed to be expensive to evaluate, and its gradient is often unavailable.
Therefore we can reduce a difficult, expensive problem to a series of
simpler, inexpensive problems – a reasonable pursuit!
Numerous acquisition functions have been proposed for Bayesian
optimization, and we will describe many popular choices in detail in
chapter 7. The most prominent means to constructing acquisition func- common Bayesian optimization policies:
tions is Bayesian decision theory, an approach to optimal decision making chapter 7, p. 123
we will discuss over the remainder of the chapter.

5.1 introduction to bayesian decision theory

Bayesian decision theory is a framework for decision making under
uncertainty that is flexible enough to handle effectively any scenario. 3 The following would be excellent companion
texts:
Instead of presenting the entire theory in complete abstraction, we will
introduce the essential concepts with an eye to the context of optimiza- m. h. degroot (1970). Optimal Statistical Deci-
tion. For a more in-depth and theoretical treatment, the interested reader sions. McGraw–Hill.
may refer to numerous comprehensive reviews of the subject.3 A good j. o. berger (1985). Statistical Decision Theory
familiarity with this material can demystify some key ideas that are often and Bayesian Analysis. Springer–Verlag.
glossed over in the Bayesian optimization literature, as it serves as the
“hidden origin” of many common acquisition functions.
In this section we will introduce to the Bayesian approach to decision
making and demonstrate how to make optimal decisions in the case of a
single isolated decision. Ultimately, we will require a theory for making
a sequence of decisions to reason over an entire optimization session. In
the next section, we will extend the line of reasoning presented below to
address sequential decision making and the construction of optimization
policies.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 89

 decision theory for optimization

Isolated decisions
A decision problem under uncertainty has two defining characteristics.
action space, A The first is the action space A, the set of all available decisions. Our
task is to select an action from this space. For example, in sequential
optimization, an optimization policy decision must select a point in the
domain X for observation, and so we have A = X.
The second critical feature is the presence of uncertain elements of
unknown variables affecting decision the world influencing the outcomes of our actions, complicating our
outcome, 𝜓 decision. Let 𝜓 represent a random variable encompassing any relevant
uncertain elements when making and evaluating a decision. Although
we may lack perfect knowledge, Bayesian inference allows us to reason
relevant observed data, D about 𝜓 in light of data via the posterior distribution 𝑝 (𝜓 | D), and we
posterior belief about 𝜓, 𝑝 (𝜓 | D) will use this belief to inform our decision.
Suppose now we must select a decision from an action space A under
uncertainty in 𝜓, informed by a set of observed data D. To guide our
utility function, 𝑢 (𝑎,𝜓, D) choice, we select a real-valued utility function 𝑢 (𝑎,𝜓, D). This function
measures the quality of selecting the action 𝑎 if the true state of the world
were revealed to be 𝜓, with higher utilities indicating more favorable
4 Typical presentations of Bayesian decision the- outcomes. The arguments to a utility function comprise everything
ory omit the data from the utility function, but required to judge the quality of a decision in hindsight: the proposed
including it offers more generality, and this
allowance will be important when we turn our action 𝑎, what we know (the data D), and what we don’t know (the
attention to optimization policies. uncertain elements 𝜓 ).4
We cannot know the exact utility that would result from selecting
any given action a priori, due to our incomplete knowledge of 𝜓. We can,
however, compute the expected utility that would result from selecting
expected utility an action 𝑎, according to our posterior belief:
∫
 
𝔼 𝑢 (𝑎,𝜓, D) | 𝑎, D = 𝑢 (𝑎,𝜓, D) 𝑝 (𝜓 | D) d𝜓. (5.2)

This expected utility maps each available action to a real value, inducing
a total order and providing a straightforward mechanism for making our
decision. We pick an action maximizing the expected utility:
 
𝑎 ∈ arg max 𝔼 𝑢 (𝑎 ,0 𝜓, D) | 𝑎 ,0 D . (5.3)
𝑎0 ∈A

5 One may question whether this framework is This decision is optimal in the sense that no other action results in greater
complete in some sense: is it possible make expected utility. (By definition!) This procedure for acting optimally
rational decisions in some other manner? The
von Neumann–Morgenstern theorem shows
under uncertainty – computing expected utility with respect to relevant
that the answer is, surprisingly, no. Assum- unknown variables and maximizing to select an action – is the central
ing a certain set of rationality axioms, any tenant of Bayesian decision making.5
rational preferences over uncertain outcomes
can be captured by the expectation of some
utility function. Thus every rational decision Example: recommending a point for use after optimization
maximizes an expected utility:
With this abstract decision-making framework established, let us analyze
j. von neumann and o. morgenstern (1944).
Theory of Games and Economic Behavior. an example decision that might be faced in the context of optimization.
Princeton University Press. [appendix a] Consider a scenario where the purpose of optimization is to identify a
single point 𝑥 ∈ X for perpetual use in a production system, preferring

90
 5.2. seqential decisions with a fixed budget

locations achieving higher values of the objective function. If we run an

optimizer and it returns some dataset D, which point should we select
for our final recommendation?
We may model this choice as a decision problem with action space
A = X, where we must reason under uncertainty about the objective
function 𝑓. We first select a utility function quantifying the quality of a
given recommendation 𝑥 in hindsight. One natural choice would be

𝑢 (𝑥, 𝑓 ) = 𝑓 (𝑥) = 𝜙,

which rewards points for achieving high values of the objective function.
Now if our optimization procedure returned a dataset D, the expected
utility from recommending a point 𝑥 is simply the posterior mean of the
recommendation
corresponding function value:
  Optimal terminal recommendation. Above:
𝔼 𝑢 (𝑥, 𝑓 ) | 𝑥, D = 𝔼[𝜙 | 𝑥, D] = 𝜇D (𝑥). (5.4) posterior belief about an objective function
given the data returned by an optimizer,
Therefore, an optimal recommendation maximizes the posterior mean: 𝑝 (𝑓 | D). Below: the expected utility for our
example, the posterior mean 𝜇 D (𝑥). The
optimal recommendation maximizes the
𝑥 ∈ arg max 𝜇D (𝑥 0).
𝑥 0 ∈X
expected utility.

Of course, other considerations in a given scenario such as risk aversion

might suggest some other utility function or action space would be more
appropriate, in which case we are free to select any alternative as we
see fit. We will discuss terminal recommendations at length in the next terminal recommendations: § 6.1, p. 109
chapter, including alternative utility functions and action spaces.

5.2 seqential decisions with a fixed budget

We have now introduced Bayesian decision theory as a framework for
computing optimal decisions informed by data. The key idea is to mea-
sure the post hoc quality of a decision with an appropriately designed
utility function, then choose actions maximizing expected utility accord-
ing to our beliefs. We will now apply this idea to the construction of
optimization policies. This setting is considerably more complicated be-
cause each decision we make over the course of optimization will shape
the context of all future decisions.

Modeling policy decisions

To define an optimization routine, we must design a policy to adaptively
design a sequence of observations seeking the optimum. Following our
discussion in the previous section, we will model each of these choices
as a decision problem under uncertainty. Some aspects of this modeling
will be straightforward and others will take some care. To begin, the
action space of each decision is the domain X, and we must act under
uncertainty about the objective function 𝑓, which induces uncertainty
about the outcomes of proposed observations. Fortunately, we may make
each decision guided by any data obtained from previous decisions.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 91

 decision theory for optimization

Bayesian inference of the objective function: To reason about uncertainty in the objective function, we follow
§ 1.2, p. 8 the path laid out in the preceding chapters and maintain a probabilistic
belief throughout optimization, 𝑝 (𝑓 | D). We make no assumptions
regarding the nature of this distribution, and in particular it need not
be a Gaussian process. Equipped with this belief, we may reason about
the result of making an observation at some point 𝑥 via the posterior
predictive distribution 𝑝 (𝑦 | 𝑥, D) (1.7), which will play a key role below.
The ultimate purpose of optimization is to collect and return a dataset
D. Before we can reason about what data we should acquire, we must
first clarify what data we would like to acquire. Following the previous
optimization utility function, 𝑢 (D) section, we will accomplish this by defining a utility function 𝑢 (D) to
evaluate the quality of data returned by an optimizer. This utility function
will serve to establish preferences over optimization outcomes: all other
things being equal, we would prefer to return a dataset with higher
utility than any dataset with lower utility. As before, we will use this
utility to guide the design of policies, by making observations that, in
utility functions for optimization: chapter 6, expectation, promise the biggest improvement in utility. We will define
p. 109 and motivate several utility functions used for optimization in the next
chapter, and some readers may wish to jump ahead to that discussion for
explicit examples before continuing. In the following, we will develop
the general theory in terms of an arbitrary utility function.

Uncertainty faced during optimization

Suppose D is a dataset of previous observations and that we must select
the next observation location 𝑥. This is the core decision defining an
optimization policy, and we will make all such decisions in the same
manner: by maximizing the expected utility of the data we will return.
Although this sounds straightforward, let us consider the uncertainty
faced when contemplating this decision in more detail. When evaluat-
ing a potential action 𝑥, uncertainty in the objective function induces
uncertainty in the corresponding value 𝑦 we will observe. Bayesian infer-
ence allows us to reason about this uncertain outcome via the posterior
predictive distribution (1.7), and we may hope to be able to address this
uncertainty without much trouble. However, we must also consider that
evaluating at 𝑥 would add the unknown observation (𝑥, 𝑦) to our dataset,
and that the contents of this updated dataset would be consulted for all
future decisions. Thus we must reason not only about the outcome of
the present observation but also its impact on the entire remainder of
optimization. This requires special attention and distinguishes sequential
decisions from the isolated decisions discussed in the last section.
Intuitively, we might suspect that decisions made closer to termina-
tion should be easier, as fewer future decisions depend on their outcomes.
This is indeed the case, and it will be prudent to define optimization poli-
6 In fact, we have already begun by analyzing a cies in reverse.6 We will first reason about the final decision – when we
decision after optimization has completed! are freed from the burden of having to ponder any future observations –
and proceed backwards to the choice of the first observation location,
working out optimal behavior every step along the way.

92
 5.2. seqential decisions with a fixed budget

In this section we will consider the construction of optimization fixed, known budget
policies assuming that we have a fixed and known budget on the number
of observations we will make. This scenario is both common in practice
and convenient for analysis, as we can for now ignore the question of
when to terminate optimization. Note that this assumption effectively
implies that every observation has a constant acquisition cost, which
may not always be reasonable. We will address variable observation cost-aware optimization: § 5.4, p. 103
costs and the question of when to stop optimization later in this chapter.
Assuming a fixed observation budget allows us to reason about opti-
mization policies in terms of the number of observations remaining to
termination, which will always be known. The problem we will consider
in this section then becomes the following: provided an arbitrary set of
data, how should we design our next evaluation location when exactly 𝜏 number of remaining observations (horizon),
observations remain before termination? In sequential decision making, 𝜏
this value is known as the decision horizon, as it indicates how far we
must look ahead into the future when reasoning about the present.
To facilitate our discussion, we pause to define notation for future
data that will be encountered during optimization relative to the present.
When considering an observation at some point 𝑥, we will call the value
resulting from an observation there 𝑦. We will then call  the dataset
available at the next stage of optimization D1 = D ∪ (𝑥, 𝑦) , where putative next observation and dataset: (𝑥, 𝑦),
the subscript indicates the number of future observations incorporated D1
into the current data. We will write (𝑥 2, 𝑦2 ) for the following observa- putative following observation and dataset:
tion, which when acquired will form D2, etc. Our final observation 𝜏 (𝑥 2 , 𝑦2 ), D2
steps in the future will then be (𝑥𝜏 , 𝑦𝜏 ), and the dataset returned by our putative final observation and dataset:
optimization procedure will be D𝜏 , with utility 𝑢 (D𝜏 ). (𝑥𝜏 , 𝑦𝜏 ), D𝜏
This utility of the data we return is our ultimate concern and will
serve as the utility function used to design every observation. Note we
may write this utility in the same form we introduced in our general
discussion:
𝑢 (D𝜏 ) = 𝑢 ( D, 𝑥, 𝑦, 𝑥 2, 𝑦2, . . . , 𝑥𝜏 , 𝑦𝜏 ),
known action unknown

which expresses the terminal utility in terms of a proposed current

action 𝑥, the known data D, and the unknown future data to be obtained:
the not-yet observed value 𝑦, and the locations {𝑥 2, . . . , 𝑥𝜏 } and values
{𝑦2, . . . , 𝑦𝜏 } of any following observations.
 
Following our treatment of isolated decisions, we evaluate a potential expected terminal utility, 𝔼 𝑢 (D𝜏 ) | 𝑥, D
observation location 𝑥 via the expected utility at termination ultimately
obtained if we observe at that point next:
 
𝔼 𝑢 (D𝜏 ) | 𝑥, D , (5.5)
and define an optimization policy via maximization:
 
𝑥 ∈ arg max 𝔼 𝑢 (D𝜏 ) | 𝑥 ,0 D . (5.6)
𝑥 0 ∈X

On its surface, this proposal is relatively simple. However, we must

now consider how to actually compute the expected terminal utility (5.5).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 93

 decision theory for optimization

Explicitly writing out the expectation over the future data in (5.5) yields
the following expression:
∫ ∫ Ö
𝜏

· · · 𝑢 (D𝜏 ) 𝑝 (𝑦 | 𝑥, D) 𝑝 (𝑥𝑖 , 𝑦𝑖 | D𝑖−1 ) d𝑦 d (𝑥𝑖 , 𝑦𝑖 ) . (5.7)
𝑖=2

This integral certainly appears unwieldy! In particular, it is unclear how

to reason about uncertainty in our future actions, as we should hope that
these actions are made to maximize our welfare rather than generated
by a random process. We will show how to compute this expression
7 This is known as bellman’s principle of opti- under the bold but rational assumption that we make all future decisions
mality, and will be discussed further later in optimally,7 and this analysis will reveal the optimal optimization policy.
this section.
We will proceed via induction on the number of evaluations remain-
ing before termination, 𝜏. We will first determine optimal behavior when
only one observation remains and then inductively consider increasingly
8 This procedure is often called “backward in- long horizons.8 For this analysis it will be useful to introduce notation
duction,” where we consider the last decision for the expected increase in utility achieved when beginning from an
first and work backward in time. Our approach
arbitrary dataset D, making an observation at 𝑥, and then continuing
of a forward induction on the horizon is equiv-
alent. optimally until termination 𝜏 steps in the future. We will write
 
𝛼𝜏 (𝑥; D) = 𝔼 𝑢 (D𝜏 ) | 𝑥, D − 𝑢 (D)

for this quantity, which is simply the expected terminal utility (5.5) shifted
by the utility of our existing data, 𝑢 (D). It is no coincidence this notation
defining a policy by maximizing an echoes our notation for acquisition functions! We will characterize the
acquisition function: § 5, p. 88 optimal optimization policy by a family of acquisition functions defined
in this manner.

Fixed budget: one observation remaining

We first consider the case where only one observation remains before
termination; that is, the horizon is 𝜏 = 1. In this case the terminal
dataset will be the current dataset augmented with a single additional
observation. As there are no following decisions to consider, we may
analyze the decision using the framework we have already developed for
isolated decisions: § 5.1, p. 89 isolated decisions. The marginal gain in utility from a final evaluation at
𝑥 is an expectation over the corresponding value 𝑦 with respect to the
posterior predictive distribution:
∫
𝛼 1 (𝑥; D) = 𝑢 (D1 ) 𝑝 (𝑦 | 𝑥, D) d𝑦 − 𝑢 (D). (5.8)

The optimal observation maximizes the expected marginal gain:

𝑥 ∈ arg max 𝛼 1 (𝑥 0; D), (5.9)

𝑥 0 ∈X

and leads to our returning a dataset with expected utility

𝑢 (D) + 𝛼 1∗ (D); 𝛼 1∗ (D) = max

0
𝛼 1 (𝑥 0; D). (5.10)
𝑥 ∈X

94
 5.2. seqential decisions with a fixed budget

observations posterior mean posterior 95% credible interval

expected marginal gain, 𝛼 1 next observation location, 𝑥 𝛼 1∗

𝑥0 𝑥

Figure 5.1: Illustration of the optimal optimization policy with

𝑝 (𝑦 | 𝑥, D) a horizon of one. Above: we compute the expected
marginal gain 𝛼 1 over the domain and design our
𝑝 (𝑦 | 𝑥 ,0 D)
next observation 𝑥 by maximizing this score. Left:
𝑢 (D1 ) − 𝑢 (D)
the computation of the expected marginal gain for
the optimal point 𝑥 and a suboptimal point 𝑥 0 indi-
cated above. In this example the marginal gain is
a simple piecewise linear function of the observed
𝑦 value (5.11), and the optimal point maximizes its
expectation.
𝑢 (D)

Here we have defined the symbol 𝛼𝜏∗ (D) to represent the expected
increase in utility when starting with D and continuing optimally for value of D with horizon 𝜏, 𝛼𝜏∗ (D)
𝜏 additional observations. This is called the value of the dataset with a
horizon of 𝜏 and will serve a central role below. We have now shown
how to compute the value of any dataset with a horizon of 𝜏 = 1 (5.10)
and how to identify a corresponding optimal action (5.9). This completes
the base case of our argument.
We illustrate the optimal optimization policy with one observation illustration of one-step optimal optimization
remaining in figure 5.1. In this scenario the belief over the objective policy
function 𝑝 (𝑓 | D) is a Gaussian process, and for simplicity we assume
our observations reveal exact values of the objective. We consider an
intuitive utility function: the maximal objective value contained in the
data, 𝑢 (D) = max 𝑓 (x).9 The marginal gain in utility offered by a putative 9 This is a special case of the simple reward util-
final observation (𝑥, 𝑦) is then a piecewise linear function of the observed ity function, which we discuss further in the
next chapter (§ 6.1, p. 109). The corresponding
value:
 expected marginal gain is the well-known ex-
𝑢 (D1 ) − 𝑢 (D) = max 𝑦 − 𝑢 (D), 0 ; (5.11) pected improvement acquisition function (§ 7.3,
p. 127).
that is, the utility increases linearly if we exceed the previously best-
seen value and otherwise remains constant. To design the optimal final
observation, we compute the expectation of this quantity over the domain
and choose the point maximizing it, as shown in the top panels. We also
illustrate the computation of this expectation for the optimal choice and

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 95

decision theory for optimization

observations posterior mean posterior 95% credible interval

expected marginal gain, 𝛼 2 next observation location, 𝑥 𝛼 2∗

Figure 5.2: Illustration of the optimal optimization policy with

a horizon of two. Above: the expected two-step
marginal gain 𝛼 2 . Right: computation of 𝛼 2 for
the optimal point 𝑥. The marginal gain is decom-
posed into two components (5.13): the immediate
gain 𝑝 (𝑦 | 𝑥, D)
 𝑢 (D1 ) − 𝑢 (D) and the expected future gain 𝑢 (D1 ) − 𝑢 (D)
𝔼 𝑢 (D2 ) − 𝑢 (D1 ) . The chosen point offers a high  
expected future reward even if the immediate re- 𝔼 𝑢 (D2 ) − 𝑢 (D1 ) 𝑦
ward is zero; see the facing page for the scenarios
resulting from the marked values. 𝑦0 𝑢 (D) 𝑦 00

a suboptimal alternative in the bottom panel. We expect an observation

at the chosen location to improve utility by a greater amount than any
alternative.

Fixed budget: two observations remaining

Rather than proceeding immediately to the inductive case, let us consider
the specific case of two observations remaining: 𝜏 = 2. Suppose we have
obtained an arbitrary dataset D and must decide where to make the
penultimate observation 𝑥. The reasoning for this special case presents
the inductive argument most clearly.
We again consider the expected increase in utility by termination,
now after two observations:
 
𝛼 2 (𝑥; D) = 𝔼 𝑢 (D2 ) | 𝑥, D − 𝑢 (D). (5.12)

Nominally this expectation requires marginalizing the observation 𝑦, as

well as the final observation location 𝑥 2 and its value 𝑦2 (5.7). However, if
we assume optimal future behavior, we can simplify our treatment of the
final decision 𝑥 2 . First we rewrite the two-step expected gain 𝛼 2 in terms
of the one-step expected gain 𝛼 1 , a function for which we have already
established a good understanding. We write the two-step difference in

96
 5.2. seqential decisions with a fixed budget



 Figure 5.3: The posterior of the objec-

 tive function given two pos-

 sible observations resulting
 𝑦 = 𝑦0
 from the optimal two-step ob-
𝛼 1∗ 
 servation 𝑥 illustrated on the
 facing page. The relatively

 low value 𝑦 0 offers no imme-

𝑥2  diate reward, but reveals a
new local optimum and the
 expected future reward from

 the optimal final decision 𝑥 2

 is high. The relatively high

 value 𝑦 00 offers a large imme-
 𝑦 = 𝑦 00 diate reward and respectable

 prospects from the optimal fi-

 nal decision as well.

𝛼 1∗


𝑥2 

utility as a telescoping sum:

   
𝑢 (D2 ) − 𝑢 (D) = 𝑢 (D1 ) − 𝑢 (D) + 𝑢 (D2 ) − 𝑢 (D1 ) ,

which yields
 
𝛼 2 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼 1 (𝑥 2 ; D1 ) | 𝑥, D .

That is, the expected increase in utility after two observations can be decomposition of expected marginal gain
decomposed as the expected increase after our first observation 𝑥 – the
expected immediate gain – plus the expected additional increase from
the final observation 𝑥 2 – the expected future gain.
It is still not clear how to address the second term in this expression.
However, from our analysis of the base case, we can reason as follows.
Given 𝑦 (and thus knowledge of D1 ), the optimal final decision 𝑥 2 (5.9)
results in an expected marginal gain of 𝛼 1∗ (D1 ), a quantity we know how
to compute (5.10). Therefore, assuming optimal future behavior, we have:
 
𝛼 2 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼 1∗ (D1 ) | 𝑥, D , (5.13)

which expresses the desired quantity as an expectation with respect

to the current observation 𝑦 only – the future value 𝛼 1∗ (5.10) does not
depend on either 𝑥 2 (due to maximization) or 𝑦2 (due to expectation).
The optimal penultimate observation location maximizes the expected
gain as usual:
𝑥 ∈ arg max 𝛼 2 (𝑥 0; D). (5.14)
𝑥 0 ∈X

and provides an expected terminal utility of

𝑢 (D) + 𝛼 2∗ (D); 𝛼 2∗ (D) = max

0
𝛼 2 (𝑥 0; D).
𝑥 ∈X

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 97

 decision theory for optimization

This demonstrates we can achieve optimal behavior for a horizon of

𝜏 = 2 and compute the value of any dataset with this horizon.
illustration of two-step optimal optimization The optimal policy with two observations remaining is illustrated in
policy figures 5.2 and 5.3. The former shows the expected two-step marginal
gain 𝛼 2 and the optimal action. This quantity depends both on the imme-
diate gain from the next observation and the expected future gain from
the optimal final action. The chosen observation appears quite promis-
ing: even if the result offers no immediate gain, it will likely provide
information that can be exploited with the optimal final decision 𝑥 2 . We
show the situation that would be faced in the final stage of optimization
for two potential values in figure 5.3. The relatively low value 𝑦 0 offers no
immediate gain but sets up an encouraging final decision, whereas the
relatively high value 𝑦 00 offers a significant immediate gain with some
chance of further improvement.

Fixed budget: inductive case

We now present the general inductive argument, which closely follows
the 𝜏 = 2 analysis above. Let 𝜏 be an arbitrary decision horizon, and for
the sake of induction assume we can compute the value of any dataset
with a horizon of 𝜏 − 1. Suppose we have an arbitrary dataset D and
must decide where to make the next observation. We will show how to
do so optimally and how to compute its value with a horizon of 𝜏.
Consider the 𝜏-step expected gain in utility from observing at some
point 𝑥:  
𝛼𝜏 (𝑥; D) = 𝔼 𝑢 (D𝜏 ) | 𝑥, D − 𝑢 (D),
10 Namely: which we seek to maximize. We decompose this expression in terms of
shorter-horizon quantities through a telescoping sum:10
𝑢 (D𝜏 ) − 𝑢 (D) =
     
𝑢 (D1 ) − 𝑢 (D) + 𝑢 (D𝜏 ) − 𝑢 (D1 ) . 𝛼𝜏 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼𝜏−1 (𝑥 2 ; D1 ) | 𝑥, D .

Now if we knew 𝑦 (and thus D1 ), optimal continued behavior would

provide an expected further gain of 𝛼𝜏−1
∗
(D1 ), a quantity we can compute
via the inductive hypothesis. Therefore, assuming optimal behavior for
all remaining decisions, we have:
 ∗ 
𝛼𝜏 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼𝜏−1 (D1 ) | 𝑥, D , (5.15)

which is an expectation with respect to 𝑦 of a function we can com-

pute. To find the optimal decision and the 𝜏-step value of the data, we
maximize:
𝑥 ∈ arg max 𝛼𝜏 (𝑥 0; D); (5.16)
𝑥 0 ∈X
𝛼𝜏∗ (D) = max 𝛼𝜏 (𝑥 0; D). (5.17)
𝑥 0 ∈X

This demonstrates we can achieve optimal behavior for a horizon of 𝜏

given an arbitrary dataset and compute its corresponding value, estab-
lishing the inductive case and completing our analysis.

98
 5.3. cost and approximation of the optimal policy

We pause to note that the value of any dataset with null horizon optimal policy: compact notation
is 𝛼 0∗ (D) = 0, and thus the expressions in (5.15–5.17) are valid for any
horizon and compactly express the proposed policy. Further, we have
actually shown that this policy is optimal in the sense of maximizing
expected terminal utility over the space of all policies, at least with optimality
respect to our model of the objective function and observations. This
follows from our induction: the base case is established in (5.9), and the 11 Since ties in (5.16) may be broken arbitrarily,
inductive case by the sequential maximization in (5.16).11 this argument does not rule out the possibility
of there being multiple, equally good optimal
policies.
Bellman optimality and the Bellman equation
Substituting (5.15) into (5.17), we may derive the following recursive 12 r. bellman (1952). On the Theory of Dy-
definition of the value in terms of the value of future data: namic Programming. Proceedings of the Na-
n  ∗ o tional Academy of Sciences 38(8):716–719.
𝛼𝜏∗ (D) = max
0
𝛼 1 (𝑥 0
; D) + 𝔼 𝛼 𝜏−1 (D 1 ) | 𝑥 0
, D . (5.18)
𝑥 ∈X

This is known as the Bellman equation and is a central result in the theory Bellman equation
of optimal sequential decisions.12 The treatment of future decisions in
this equation – recursively assuming that we will always act to maximize
expected terminal utility given the available data – reflects bellman’s bellman’s principle of optimality
principle of optimality, which characterizes optimal sequential decision
policies in terms of the optimality of subpolicies:13
An optimal policy has the property that whatever the initial state 13 r. bellman (1957). Dynamic Programming.
and initial decision are, the remaining decisions must constitute Princeton University Press.
an optimal policy with regard to the state resulting from the first
decision.
That is, to make a sequence of optimal decisions, we make the first
decision optimally, then make all following decisions optimally given
the outcome!

5.3 cost and approximation of the optimal policy

Although the framework presented in the previous section is concep-
tually simple and theoretically attractive, the optimal policy is unfortu-
nately prohibitive to compute except for very short decision horizons.
To demonstrate the key computational barrier, consider the selection
of the penultimate observation location. The expected two-step marginal
gain to be maximized is (5.13):
 
𝛼 2 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼 1∗ (D1 ) | 𝑥, D .
The second term appears to be a straightforward expectation over the
one-dimensional random variable 𝑦. However, evaluating the integrand
in this expectation requires solving a nontrivial global optimization
problem (5.10)! Even with only two evaluations remaining, we must
solve a doubly nested global optimization problem, an onerous task.
Close inspection of the recursively defined optimal policy (5.15–5.16) “unrolling” the optimal sequential policy
reveals that when faced with a horizon of 𝜏, we must solve 𝜏 nested

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 99

 decision theory for optimization

𝑦 𝑦2 𝑥𝜏 𝑢 (D𝜏 )
𝑥 𝑥2 ··· 𝑦𝜏

arg max 𝔼𝑦 max 𝔼𝑦2 ··· max 𝔼𝑦𝜏 𝑢 (D𝜏 )

𝑥 𝑥2 𝑥𝜏

Figure 5.4: The optimal optimization policy as a decision tree. Squares indicate decisions (the choice of each observation),
and circles represent expectations with respect to random variables (the outcomes of observations). Only one
possible optimization path is shown; dangling edges lead to different futures, and all possibilities are always
considered. We maximize the expected terminal utility 𝑢 (D𝜏 ), recursively assuming optimal future behavior.

optimization problems to find the optimal decision. Temporarily adopting

compact notation, we may “unroll” the optimal policy as follows:

𝑥 ∈ arg max 𝛼𝜏 ;
∗
𝛼𝜏 = 𝛼 1 + 𝔼[𝛼𝜏−1 ]
= 𝛼 1 + 𝔼[max 𝛼𝜏−1 ]
  
= 𝛼 1 + 𝔼 max 𝛼 1 + 𝔼[𝛼𝜏−2
∗
]
h n   i
= 𝛼 1 + 𝔼 max 𝛼 1 + 𝔼 max 𝛼 1 + 𝔼[max{𝛼 1 + · · · .

The design of each optimal decision requires repeated maximization

over the domain and expectation over unknown observations until the
horizon is reached. This computation is visualized as a decision tree in
figure 5.4, where it is clear that each unknown quantity contributes a
significant branching factor. Computing the expected utility at 𝑥 exactly
requires a complete traversal of this tree.
running time of optimal policy The cost of computing the optimal policy clearly grows with the
horizon. Let us perform a careful running time analysis for a naïve
implementation via exhaustive traversal of the decision tree in figure 5.4
with off-the-shelf procedures. Suppose we use an optimization routine
for each maximization and a numerical quadrature routine for each
expectation encountered in this computation. If we allow 𝑛 evaluations
evaluation budget for optimization, 𝑛 of the objective for each call to the optimizer and 𝑞 observations of the
evaluation budget for quadrature, 𝑞 integrand for each call to the quadrature routine, then each decision
along the horizon will contribute a multiplicative factor of O(𝑛𝑞) to the
total running time. Computing the optimal decision with a horizon of
𝜏 thus requires O(𝑛𝜏𝑞𝜏 ) work, an exponential growth in running time
14 Detailed references are provided by: with respect to the horizon.
w. b. powell (2011). Approximate Dynamic Pro- Evidently, the computational effort required for realizing the optimal
gramming: Solving the Curses of Dimensional- policy quickly becomes intractable, and we must find some alternative
ity. John Wiley & Sons. mechanism for designing effective optimization policies. General approx-
d. p. bertsekas (2017). Dynamic Programming imation schemes for the optimal policy have been studied in depth under
and Optimal Control. Vol. 1. Athena Scientific. the name approximate dynamic programming,14 and usually operate as

100
 5.3. cost and approximation of the optimal policy

Figure 5.5: A lookahead approximation

to the optimal optimization
𝑦 𝑥ℓ 𝑢 (Dℓ ) policy. We choose the optimal
𝑥 ··· 𝑦ℓ
decision for a limited horizon
ℓ  𝜏 decisions, ignoring any
observations that would fol-
low.

follows. We begin with the intractable optimal expected marginal gain

(5.15):  ∗ 
𝛼𝜏 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼𝜏−1 (D1 ) | 𝑥, D ,
and substitute a tractable approximation for the “hard” part of the ex-
pression: the recursively defined future value 𝛼 ∗ (5.18). The result is an
acquisition function inducing a suboptimal – but rationally guided – ap-
proximate policy. Two particular approximations schemes have proven
useful in Bayesian optimization: limited lookahead and rollout.

Limited lookahead
One widespread and surprisingly effective approximation is to simply
limit how many future observations we consider in each decision. This
is practical as decisions closer to termination require substantially less
computation than earlier decisions.
With this in mind, we can construct a natural family of approxima-
tions to the optimal policy defined by artificially limiting the horizon
used throughout optimization to some computationally feasible maxi-
mum ℓ. When faced with an infeasible decision horizon 𝜏, we make the
crude approximation

𝛼𝜏 (𝑥; D) ≈ 𝛼 ℓ (𝑥; D),

and by maximizing this score, we act optimally under the incorrect but
convenient assumption that only ℓ observations remain. This effectively 15 Equivalently, we approximate the true future
assumes 𝑢 (D𝜏 ) ≈ 𝑢 (Dℓ ).15 This may be reasonable if we expect decreas- value 𝛼𝜏∗ −1 with 𝛼 ℓ∗−1 .
ing marginal gains, implying a significant fraction of potential gains can
be attained within the truncated horizon. This scheme is often described myopic approximations
(sometimes disparagingly) as myopic, as we limit our sight to only the
next few observations rather than looking ahead to the full horizon.
A policy that designs each observation to maximize the limited-
horizon acquisition function 𝛼 min{ℓ,𝜏 } is called an ℓ-step lookahead pol- ℓ-step lookahead
icy.16 This is also called a rolling horizon strategy, as the fixed horizon rolling horizon
“rolls along” with us as we go. By limiting the horizon, we bound the
computational effort required for each decision to at-most O(𝑛 ℓ 𝑞 ℓ ) time 16 We take the minimum to ensure we don’t look
with the implementation described above. This can be a considerable beyond the true horizon, which would be non-
sense.
savings when the observation budget is much greater than the selected
lookahead. A lookahead policy is illustrated as a decision tree in figure
5.5. Comparing to the optimal policy in figure 5.4, we simply “cut off”
and ignore any portion of the tree lying deeper than ℓ steps in the future.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 101

decision theory for optimization

𝑦 𝑦2 𝑥𝜏 𝑢 (D𝜏 )
𝑥 𝑥2 ··· 𝑦𝜏

Figure 5.6: A decision tree representing a rollout policy. Comparing to the optimal policy in figure 5.4, we simulate future
decisions starting with 𝑥 2 using an efficient but suboptimal heuristic policy, rather than the intractable optimal
policy. We maximize the expected terminal utility 𝑢 (D𝜏 ), assuming potentially suboptimal future behavior.

one-step lookahead Particularly important in Bayesian optimization is the special case of

one-step lookahead, which successively maximizes the expected marginal
gain after acquiring a single additional observation, 𝛼 1 . One-step looka-
head is the most efficient lookahead approximation (barring the absurdity
that would be “zero-step” lookahead), and it is often possible to derive
closed-form, analytically differentiable expressions for 𝛼 1 , enabling effi-
cient implementation. Many well-known acquisition functions represent
common Bayesian optimization policies: one-step lookahead approximations for some implicit choice of utility
chapter 7, p. 123 function, as we will see in chapter 7.

Rollout
The optimal policy evaluates a potential observation location by simu-
lating the entire remainder of optimization following that choice, recur-
sively assuming we will use the optimal policy for every future decision.
Although sensible, this is clearly intractable. Rollout is an approach to ap-
proximate policy design that emulates the structure of the optimal policy,
but using a tractable suboptimal policy to simulate future decisions.
A rollout policy is illustrated as a decision tree in figure 5.6. Given a
base policy, heuristic policy putative next observation (𝑥, 𝑦), we use an inexpensive so-called base
or heuristic policy to simulate a plausible – but perhaps suboptimal –
realization of the following decision 𝑥 2 . Note there is no branching in
the tree corresponding to this decision, as it does not depend on the
exhaustively enumerated subtree required by the optimal policy. We
then take an expectation with respect to the unknown value 𝑦2 as usual.
Given a putative value of 𝑦2 , we use the base policy to select 𝑥 3 and
continue in this manner until reaching the decision horizon. We use the
terminal utilities in the resulting pruned tree to estimate the expected
marginal gain 𝛼𝜏 , which we maximize as a function of 𝑥.
choice of base policy There are no constraints on the design of the base policy used in
rollout; however, for this approximation to be sensible, we must choose
something relatively efficient. One common and often effective choice
is to simulate future decisions with one-step lookahead. If we again
use off-the-shelf optimization and quadrature routines to traverse the
rollout decision tree in figure 5.6 with this particular choice, the running
time of the policy with a horizon of 𝜏 is O(𝑛 2𝑞𝜏 ), significantly faster

102
 5.4. cost-aware optimization and termination as a decision

Figure 5.7: A batch rollout policy as a de-


 𝑥2  
𝑦2  cision tree. Given a putative

.
 
 
.
 

𝑦 𝑢 (D𝜏 ) value for the next evaluation
.. ..
 

𝑥

  
𝑥𝜏 
(𝑥, 𝑦), we design all remain-
  
  𝑦𝜏  ing decisions simultaneously
using a batch base policy and
take the expectation of the ter-
minal utility with respect to
their values.
than the optimal policy. Although there is still exponential growth with
respect to 𝑞, we typically have 𝑞  𝑛,17 so we can usually entertain
farther horizons with rollout than with limited lookahead with the same 17 For estimating a one-dimensional expectation
amount of computational effort. we might take 𝑞 on the order of roughly 10,
but for optimizing a nonconvex acquisition
Due to the flexibility in the design of the base policy, rollout is a re- function over the domain we might take 𝑛 on
markably flexible approximation scheme. For example, we can combine the order of thousands or more.
rollout with the idea of limiting the decision horizon to yield approxi-
mate policies with tunable running time. In fact, we can interpret ℓ-step
lookahead as a special case of rollout, where the base policy designs limited lookahead as rollout
the next ℓ − 1 decisions optimally assuming a myopic horizon and then
simply terminates early, discarding any remaining budget.
We may also adopt a base policy that designs all remaining observa-
tions simultaneously. Ignoring the dependence between these decisions
can provide a computational advantage while retaining awareness of the
evolving decision horizon, and such batch rollout schemes have proven batch rollout
useful in Bayesian optimization. A batch rollout policy is illustrated as a
decision tree in figure 5.7. Although we account for the entire horizon,
the tree depth is reduced dramatically compared to the optimal policy.

5.4 cost-aware optimization and termination as a decision

Thus far we have only considered the construction of optimization poli-
cies under a known budget on the total number of observations. Although
this scenario is pervasive, it is not universal. In some situations, we might
wish instead to use our evolving beliefs about the objective function to
decide dynamically when termination is the best course of action.
Dynamic termination can be especially prudent when we want to
reason explicitly about the cost of data acquisition during optimization.
For example, if this cost were to vary across the domain, it would not
be sensible to define a budget in terms of function evaluations. How-
ever, by accounting for observation costs in the utility function, we can
reason about cost–benefit tradeoffs during optimization and seek to ter-
minate whenever the expected cost of further observation outweighs
any expected benefit it might provide.

Modeling termination decisions and the optimal policy

We consider a modification to the sequential decision problem we an-
alyzed in the known-budget case, wherein we now allow ourselves to

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 103

 decision theory for optimization

terminate optimization at any time of our choosing. Suppose we are at an

arbitrary point of optimization and have already obtained data D. We face
the following decision: should we terminate optimization immediately
and return D? If not, where should we make our next observation?
action space, A We model this scenario as a decision problem under uncertainty
with an action space equal to the domain X, representing potential
observation locations if we decide to continue, augmented with a special
termination option, ∅ additional action ∅ representing immediate termination:

A = X ∪ {∅}. (5.19)

For the sake of analysis, after the termination action has been selected, it
is convenient to model the decision process as not actually terminating,
but rather continuing with the collapsed action space A = {∅} – once
you terminate, there’s no going back.
As before, we may derive the optimal optimization policy in the
adaptive termination case via induction on the decision horizon 𝜏. How-
ever, we must address one technical issue: the base case of the induction,
which analyzes the “final” decision, breaks down if we allow the possi-
bility of a nonterminating sequence of decisions. To sidestep this issue,
bound on total number of observations, 𝜏max we assume there is a fixed and known upper bound 𝜏max on the total
number of observations we may make, at which point optimization is
compelled to terminate regardless of any other concern. This is not an
overly restrictive assumption in the context of Bayesian optimization.
Because observations are assumed to be expensive, we can adopt some
18 It is possible to consider unbounded sequen- suitably absurd upper bound without issue; for example, 𝜏max = 1 000 000
tial decision problems, but this is probably not would suffice for an overwhelming majority of plausible scenarios.18
of practical interest in Bayesian optimization:
After assuming the decision process is bounded, our previous induc-
m. h. degroot (1970). Optimal Statistical Deci- tive argument carries through after we demonstrate how to compute
sions. McGraw–Hill. [§ 12.7] the value of the termination action. Fortunately, this is straightforward:
termination does not augment our data, and once this action is taken, no
other action will ever again be allowed. Therefore the expected marginal
gain from termination is always zero:

𝛼𝜏 (∅; D) = 0. (5.20)

With this, substituting A for X in (5.15–5.17) now gives the optimal policy.
Intuitively, the result in (5.20) implies that termination is only the
optimal decision if there is no observation offering positive expected
gain in utility. For the utility functions described in the next chapter –
all of which are agnostic to costs and measure optimization progress
19 This can be proven through various “informa- alone – reaching this state is actually impossible.19 However, explicitly
tion never hurts” (in expectation) results. accounting for observation costs in addition to optimization progress in
the utility function resolves this issue, as we will demonstrate.

Example: cost-aware optimization

To illustrate the behavior of a policy allowing early termination, we
return to our motivating scenario of accounting for observation costs.

104
 5.4. cost-aware optimization and termination as a decision

observations posterior mean posterior 95% credible interval

expected gain to data utility, 𝛼 10 observation cost, 𝑐

expected cost-adjusted gain, 𝛼 1 next observation location, 𝑥

𝛼 1∗

Figure 5.8: Illustration of one-step lookahead with the option to terminate. With a linear utility and additive costs, the
expected marginal gain 𝛼 1 is the expected marginal gain to the data utility 𝛼 10 adjusted for the cost of acquisition
𝑐. For some points, the cost-adjusted expected gain is negative, in which case we would prefer immediate
termination to observing there. However, continuing with the chosen point is expected to increase the utility
of the current data.

Consider the objective function belief in the top panel of figure 5.8 (which
is identical to that from our running example from figures 5.1–5.3) and 20 We will consider unknown and stochastic
suppose that the cost of observation now depends on location according costs in § 11.1, p. 245.
to a known cost function 𝑐 (𝑥),20 illustrated in the middle panel. observation cost function, 𝑐 (𝑥)
If we wish to reason about observation costs in the optimization
policy, we must account for them somehow, and the most natural place
to do so is in the utility function. Depending on the situation, there are
many ways we could proceed;21 however, one natural approach is to 21 We wish to stress this point – there is consider-
first select a utility function measuring the quality of a returned dataset able flexibility beyond the scheme we describe.
alone, ignoring any costs incurred to acquire it. We call this quantity
the data utility and notate it with 𝑢 0(D). The data utility is akin to the data utility, 𝑢 0 (D)
cost-agnostic utility from the known-budget case, and any one of the utility functions for optimization: chapter 6,
options described in the next chapter could reasonably fill this role. p. 109
We now adjust the data utility to account for the cost of data acquisi-
tion. In many applications, these costs are additive, so that the total cost
of gathering a dataset D is simply observation costs, 𝑐 (D)
∑︁
𝑐 (D) = 𝑐 (𝑥). (5.21)
𝑥 ∈D 22 Some additional discussion on this natural ap-
proach can be found in:
If the acquisition cost can be expressed in the same units as the data
utility – for example, if both can be expressed in monetary terms22– then h. raiffa and r. schlaifer (1961). Applied Sta-
tistical Decision Theory. Division of Research,
we might reasonably evaluate a dataset D by the cost-adjusted utility: Graduate School of Business Administration,
Harvard University. [chapter 4]
𝑢 (D) = 𝑢 0(D) − 𝑐 (D). (5.22)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 105

decision theory for optimization

Demonstration: one-step lookahead with cost-aware utility

Returning to the scenario in figure 5.8, let us adopt a cost-aware utility
function of the above form (5.22) and consider the behavior of a one-step
lookahead approximation to the optimal optimization policy.
For these choices, if we were to continue optimization by evaluating
at a point 𝑥, the resulting one-step marginal gain in utility would be:
 
𝑢 (D1 ) − 𝑢 (D) = 𝑢 0(D1 ) − 𝑢 0(D) − 𝑐 (𝑥),
the cost-adjusted marginal gain in the data utility alone. Therefore the
expected marginal gain in utility is:
𝛼 1 (𝑥; D) = 𝛼 10 (𝑥; D) − 𝑐 (𝑥),
where 𝛼 10 is the one-step expected gain in the data utility (5.8). That
is, we simply adjust what would have been the acquisition function in
the cost-agnostic setting by subtracting the cost of data acquisition. To
prefer evaluating at 𝑥 to immediate termination, this quantity must have
positive expected value (5.20).
example and discussion The resulting policy is illustrated in figure 5.8. The middle panel
shows the cost-agnostic acquisition function 𝛼 10 (from figure 5.1), which
is then adjusted for observation cost in the bottom panel. This renders the
expected marginal gain negative in some locations, where observations
are not expected to be worth their cost. However, in this case there are still
regions where observation is favored to termination, and optimization
continues at the selected location. Comparing with the cost-agnostic
setting in figure 5.1, the optimal observation has shifted from the right-
hand side to the left-hand side of the previously best-seen point, as an
observation there is more cost effective.

5.5 summary of major ideas

defining optimization policies via acquisition • Optimization policies can be conveniently defined via an acquisition
functions: p. 88 function assigning a score to each potential observation location. We
then design observations by maximizing the acquisition function (5.1).
• Bayesian decision theory is a general framework for optimal decision
making under uncertainty, through which we can derive optimal opti-
mization policies and stopping rules.
introduction to Bayesian decision theory: • The key elements of a decision problem under uncertainty are:
§ 5.1, p. 89
– an action space A, from which we must choose an action 𝑎,
– uncertainty in elements 𝜓 relevant to the decision, represented by a
posterior belief 𝑝 (𝜓 | D), and
– a utility function 𝑢 (𝑎,𝜓, D) quantifying the quality of the action 𝑎
assuming a given realization of the uncertain elements 𝜓.

Given these, an optimal decision maximizes the expected utility (5.2–5.3).

modeling policy decisions: § 5.2, p. 91 • Optimization policy decisions may be cast in this framework by defining
a utility function for the data returned by an optimizer, then designing

106
 5.5. summary of major ideas

each observation location to maximize the expected utility with respect

to all future data yet to be obtained (5.5–5.6).
• To ensure the optimality of a sequence of decisions, we must recursively bellman’s principle of optimality: § 5.2, p. 99
assume the optimality of all future decisions. This is known as bellman’s
principle of optimality. Under this assumption, the optimal policy can be
derived inductively and assumes a simple recursive form (5.15–5.17).
• The cost of computing the optimal policy grows exponentially with the computational burden and approximation of
decision horizon, but several techniques under the umbrella approximate the optimal policy: § 5.3, p. 99
dynamic programming provide tractable approximations. Two notable
examples are limited lookahead, where the decision horizon is artificially
limited, and rollout, where future decisions are simulated suboptimally.
• Through careful accounting, we may explicitly account for the (possibly termination as a decision: § 5.4, p. 103
nonuniform) cost of data acquisition in the utility function. Offering a ter-
mination option and computing the resulting optimal policy then allows
us to adaptively terminate optimization when continuing optimization
becomes a losing battle of cost versus expected gain.

In the next chapter we will discuss several prominent utility functions for
measuring the quality of a dataset returned by an optimization procedure.
In the following chapter, we will demonstrate how many common acqui- common Bayesian optimization policies:
sition functions for Bayesian optimization may be realized by performing chapter 7, p. 123
one-step lookahead with these utility functions.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 107

 6
U TILI TY F UNCTIONS FOR OPTIMIZATION

In the last chapter we introduced Bayesian decision theory, a framework

for decision making under uncertainty through which can derive theo-
retically optimal optimization policies. Central to this approach is the
notion of a utility function evaluating the quality of a dataset returned
from an optimization routine. Given a model of the objective function –
conveying our beliefs in the face of uncertainty – and a utility function
– expressing our preferences over outcomes – computing the optimal
policy is purely mechanical: we design every observation to maximize
the expected utility of the returned dataset (5.15–5.17). Setting aside com-
putational issues, adopting this approach entails only two major hurdles:
building an objective function model consistent with our beliefs and
designing a utility function consistent with our preferences.
Neither of these tasks is trivial! Beliefs and preferences are so innate
to the human experience that distilling them down to mathematical
symbols can be challenging. Fortunately, expressive and mathematically
convenient options for both are readily available. We devoted significant
attention to model building in the first part of this book, and we will
address the construction of utility functions in this chapter. We will in-
troduce a number of common utility functions designed for optimization,
each carrying a different perspective on how optimization performance
should be quantified. We hope that the underlying motivation for these
utility functions may inspire the design of novel alternatives when called
for. In the next chapter, we will demonstrate how approximating the opti- common Bayesian optimization policies: § 7,
mal optimization policy corresponding to the utility functions described p. 123
here yields many widespread Bayesian optimization algorithms.
Although we will be using Gaussian process models in our illustra-
tions throughout the chapter, we will not assume the objective function
model is a Gaussian process in our discussion. As in the previous chap-
ters, we will use the notation 𝜇D (𝑥) = 𝔼[𝜙 | 𝑥, D] for the posterior mean posterior mean function, 𝜇D
of the objective function; this should not be interpreted as implying any
particular model structure beyond admitting a posterior mean.

6.1 expected utility of terminal recommendation

The purpose of optimization is often to explore a space of possibilities in selecting a point for permanent use
search of the single best alternative, and after investing in optimization,
we commit to using some chosen point in a subsequent procedure. In
this context, the only purpose of the data collected during optimization
is to help select this final point. For example, in hyperparameter tuning,
we may evaluate numerous hyperparameters during model development,
only to use the apparently best settings found in a production system.
Selecting a point for permanent use represents a decision, which Bayesian decision theory: chapter 5, p. 87
we may analyze using Bayesian decision theory. If the sole purpose of
optimization is to inform a final decision, it is natural to design the policy
to maximize the expected utility of the terminal decision directly, and
several popular policies are defined in this manner.

This material will be published by Cambridge University Press as Bayesian Optimization. 109
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 utility functions for optimization

Formalization of terminal recommendation decision

Suppose we have run an optimization routine, which returned a dataset
D = (x, y), and suppose we now wish to recommend a point 𝑥 ∈ X
for use in some task, with performance determined by the underlying
1 Dependence on 𝜙 alone is not strictly neces- objective function value 𝜙 = 𝑓 (𝑥).1 This represents a decision under
sary. For example, in the interest of robustness uncertainty about 𝜙, informed by the predictive distribution, 𝑝 (𝜙 | 𝑥, D).
we might wish to ensure that function values
To completely specify the decision problem, we must identify an
are high in the neighborhood of our recom-
mendation was well. This would possible in action space A ⊂ X for our recommendation and a utility function 𝜐 (𝜙)
the same framework by redefining the utility evaluating a recommendation in hindsight according to its objective
function as desired. value 𝜙. Given these, a rational recommendation maximizes the expected
utility:  
𝑥 ∈ arg max 𝔼 𝜐 (𝜙 0) | 𝑥 ,0 D .
𝑥 0 ∈A
The expected utility of the optimal recommendation only depends on
the data returned by the optimizer; it does not depend on the optimal
recommendation 𝑥 (due to maximization) nor its objective value 𝜙 (due
𝑥 𝜐 (𝜙) to expectation). This suggests to a natural utility for use in optimization:
the expected quality of an optimal terminal recommendation given the
data,  
= 𝑢 (D) 𝑢 (D) = max
0
𝔼 𝜐 (𝜙 0) | 𝑥 ,0 D . (6.1)
𝑥 ∈A
We may also interpret this class of utility func- In the context of the sequential decision tree from figure 5.4, this utility
tions as augmenting the decision tree in figure
5.4 with a final layer corresponding to the ter-
function effectively “collapses” the expected utility of a final decision
minal decision. The utility of the data is then into a utility for the returned data; see the illustration in the margin.
the expected utility of this subtree, assuming We are free to select the action space and utility function for the final
optimal behavior. recommendation as we see fit; we provide some advice below.

Choosing an action space

We begin with the action space A ⊂ X. One extreme option is to restrict
our choice to only the visited points x. This ensures at least some knowl-
edge of the objective function at the recommended point, which may be
prudent when the objective function model may be misspecified. The
other extreme is the maximally permissive alternative: the entire domain
X, allowing us to recommend any point, including those arbitrarily far
from our observations. The wisdom of recommending an unvisited point
for perpetual use is ultimately a question of faith in the model’s beliefs.
bound on uncertainty Compromises between these extremes have also been occasionally
suggested in the literature. osborne et al. for example proposed restrict-
ing the choice of final recommendation to only those points where the
2 m. a. osborne et al. (2009). Gaussian Processes objective function is known with acceptable tolerance.2 Such a scheme
for Global Optimization. lion 3. can limit unwanted surprise from recommending points where the ob-
jective function value is not known with sufficient certainty. One might
accomplish this in several ways; osborne et al. adopted a parametric,
data-dependent action space of the form

A(𝜀; D) = 𝑥 | std[𝜙 | 𝑥, D] ≤ 𝜀 ,
where 𝜀 is a threshold specifying the largest acceptable uncertainty.

110
 6.1. expected utility of terminal recommendation

Choosing a utility function and risk tolerance

In addition to selecting an action space, we must also select a utility func-
tion 𝜐 (𝜙) evaluating a recommendation at 𝑥 in light of the corresponding
function value 𝜙. As our focus is on maximization (1.1), it is clear that the
utility should be monotonically increasing in 𝜙, but it is not necessarily
clear what shape this function should assume. The answer depends on 𝜙
our risk tolerance, a concept demonstrated in the margin. When mak-
ing our final recommendation, we may wish to consider not only the Consider the illustrated beliefs about the
expected function value of a given point but also our uncertainty in this objective function value corresponding to two
value, as points with greater uncertainty may result in more surprising possible recommendations. The blue option
has a higher expected value, but also greater
and potentially disappointing results. uncertainty, and proposing it entails some risk.
By controlling the shape of the utility function 𝜐 (𝜙), we may induce The red alternative has a lower expected value
different behavior with respect to risk. The simplest and most common but is perhaps a safer option. A risk-averse
option encountered in Bayesian optimization is a linear utility: agent might prefer the red point, whereas a
risk-tolerant agent might prefer the blue point.
𝜐 (𝜙) = 𝜙 . (6.2)

In this case, the expected utility from recommending 𝑥 is simply the 𝜐 (𝜙)
posterior mean of 𝜙, as we have already seen (5.4):
 
𝔼 𝜐 (𝜙) | 𝑥, D = 𝜇D (𝑥),

and an optimal recommendation maximizes the posterior mean over the

action space:
𝑥 = arg max 𝜇D (𝑥 0). 𝜙
𝑥 0 ∈A

Uncertainty in the objective function is not considered in this decision at A risk-neutral (linear) utility function.
all! Rather, we are indifferent between points with equal expected value,
regardless of their uncertainty – that is, we are risk neutral.
Risk neutrality is computationally convenient due to the simple form
of the expected utility, but may not always reflect our true preferences.
In the margin we show beliefs over the objective values for two potential
recommendations with equal expected value but significantly different
𝜙
risk. In many scenarios we would have a clear preference between the
two alternatives, but a risk-neutral utility induces complete indifference.
Beliefs over two recommendations with equal
A useful concept when reasoning about risk preferences is the so- expected value. A risk-neutral agent would
called certainty equivalent. Consider a risky potential recommendation be indifferent between these alternatives, a
𝑥, that is, a point for which we do not know the objective value exactly. risk-averse agent would prefer the red option,
and a risk-seeking agent would prefer the
The certainty equivalent for 𝑥 is the value of a hypothetical risk-free
blue option.
alternative for which our preferences would be indifferent. That is, the
certainty equivalent for 𝑥 corresponds to an objective function value 𝜙 0
such that  
𝜐 (𝜙 0) = 𝔼 𝜐 (𝜙) | 𝑥, D .
Under a risk-neutral utility function, the certainty equivalent of a point
𝑥 is simply its expected value: 𝜙 0 = 𝜇D (𝑥). Thus we would abandon a po-
tential recommendation for another only if it had greater expected value,
independent of risk. However, we may encode risk-aware preferences
with appropriately designed nonlinear utility functions.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 111

 utility functions for optimization

If our preferences indicate risk aversion, we might be willing to rec-

𝜐 (𝜙) ommend a point with lower expected value if it also entailed less risk.
We may induce risk-averse preferences by adopting a utility function
that is a concave function of the objective value. In this case, by Jensen’s
inequality we have
 


𝜐 (𝜙 0) = 𝔼 𝜐 (𝜙) | 𝑥, D ≤ 𝜐 𝔼[𝜙 | 𝑥, D] = 𝜐 𝜇D (𝑥) ,

and thus the certainty equivalent of a risky recommendation is less

𝜙
than its expected value; see the example in the margin. Similarly, we
may induce risk-seeking preferences with a convex utility function, in
A risk-averse (concave) utility function.
which case the certainty equivalent of a risky recommendation is greater
than its expected value – our preferences encode an inclination toward
gambling. Risk-averse and risk-seeking utilities are rarely encountered
𝜙 in the Bayesian optimization literature; however, they may be preferable
in some practical settings, as risk neutrality is often questionable.
A risk-averse agent may be indifferent be- Numerous risk-averse utility functions have been proposed in the
tween the risky recommendation in blue and economics and decision theory literature,3 and a full discussion is beyond
its risk-free certainty equivalent with lower the scope of this book. However, one natural approach is to quantify the
expected value in red.
risk associated with recommending an uncertain value 𝜙 by its standard
deviation:
3 One flexible family is the hyperbolic abso- 𝜎 = std[𝜙 | 𝑥, D].
lute risk aversion (hara) class, which includes
many popular choices as special cases: Now we may establish preferences over potential recommendations
consistent with4 a weighted combination of a point 𝑥’s expected reward,
j. e. ingersoll jr. (1987). Theory of Financial
Decision Making. Rowman & Littlefield. [chap- 𝜇 = 𝜇D (𝑥), and its risk, 𝜎:5
ter 1] 𝜇 + 𝛽𝜎.
4 Under Gaussian beliefs on function values, one Here 𝛽 serves as a tunable risk-tolerance parameter: values 𝛽 < 0 penal-
can find a family of concave (or convex) utility
functions inducing equivalent recommenda-
ize risk and induce risk-averse behavior, values 𝛽 > 0 reward risk and
tions. See § 8.26, p. 171 for related discussion. induce risk-seeking behavior, and 𝛽 = 0 induces risk neutrality (6.2).
5 Note that expected reward (𝜇) and risk (𝜎) Two particular utility functions from this general framework are
have compatible units in this formulation, so widely encountered in Bayesian optimization, both representing the
this weighted combination is sensible. expected utility of a risk-neutral optimal terminal recommendation.

Simple reward
Suppose an optimization routine returned data D = (x, y) to inform a
terminal recommendation, and that we will make this decision using
the risk-neutral utility function 𝜐 (𝜙) = 𝜙 (6.2). If we limit the action
space of this recommendation to only the locations evaluated during
optimization x, the expected utility of the optimal recommendation is
6 This name contrasts with the cumulative re- the so-called simple reward:6, 7
ward: § 6.2, p. 114.
7 One technical caveat is in order: when the 𝑢 (D) = max 𝜇D (x). (6.3)
dataset is empty, the maximum degererates
and we have 𝑢 (∅) = −∞. In the special case of exact observations, where y = 𝑓 (x) = 𝝓, the
simple reward reduces to the maximal function value encountered during
optimization:
𝑢 (D) = max 𝝓. (6.4)

112
 6.1. expected utility of terminal recommendation

global reward
simple reward Figure 6.1: The terminal recommenda-
tions corresponding to the
simple reward and global re-
ward for an example dataset
comprising five observations.
The prior distribution for the
objective for this demonstra-
tion is illustrated in figure 6.3.

recommendations

One-step lookahead with the simple reward utility function pro- expected improvement: § 7.3 p. 127
duces a widely-used acquisition function known as expected improvement,
which we will discuss in detail in the next two chapters.

Global reward
Another prominent utility is the global reward.8 Here we again consider a 8 “Global simple reward” would be a more accu-
risk-neutral terminal recommendation, but now expand the action space rate (but annoyingly bulky) name.
for this recommendation to the entire domain X. The expected utility of
this recommendation is the global maximum of the posterior mean:

𝑢 (D) = max 𝜇D (𝑥). (6.5)

𝑥 ∈X

An example dataset exhibiting a large discrepancy between the simple

reward (6.3) and global reward (6.5) utilities is illustrated in figure 6.1.
The larger action space underlying global reward leads to a markedly
different and somewhat riskier recommendation.
One-step lookahead with global reward (6.5) yields the knowledge knowledge gradient: § 7.4 p. 129
gradient acquisition function, which we will also consider at length in
the following chapters.

A tempting, but nonsensical alternative

There is an alternative utility deceptively similar to the simple reward
that is sometimes encountered in the Bayesian optimization literature,
?
namely the maximum noisy observed value contained in the dataset:9 9 The “questionable equality” symbol = is re-
served for this single dubious equation.
?
𝑢 (D) = max y. (6.6)

In the case of exact observations of the objective function, this value

coincides with the simple reward (6.4), which has a natural interpretation
as the expected utility of a particular optimal terminal recommendation.
However, this correspondence does not hold in the case of inexact or
noisy observations, and the proposed utility is rendered absurd.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 113

utility functions for optimization

max y

Figure 6.2: The utility 𝑢 (D) = max y would prefer the excessively noisy dataset on the left to the less-noisy dataset
on the right with smaller maximum value. The data on the left reveal little information about the objective
function, and the maximum observed value is very likely to be an outlier, whereas the data on the right indicate
reasonable progress.

large-but-noisy observations are not This is simple to demonstrate by contemplating the preferences over
necessarily preferable outcomes encoded in the utility, which may not align with intuition.
This disparity is especially notable in situations with excessively noisy
observations, where the maximum value observed will likely reflect
spurious noise rather than actual optimization progress.
example and discussion Figure 6.2 shows an extreme but illustrative example. We consider
two optimization outcomes over the same domain, one with excessively
noisy observations and the other with exact measurements. The noisy
dataset contains a large observation on the right-hand side of the domain,
but this is almost certainly the result of noise, as indicated by the objective
function posterior. Although the other dataset has a lower maximal value,
the observations are more trustworthy and represent a plainly better
outcome. But the proposed utility (6.6) prefers the noisier dataset! On
the other hand, both the simple and global reward utilities prefer the
noiseless dataset, as the data produce a larger effect on the posterior
mean – and thus yield more promising recommendations.
approximation to the simple reward Of course, errors in noisy measurements are not always as extreme
as in this example. When the signal-to-noise ratio is relatively high, the
utility (6.6) can serve as a reasonable approximation to the simple reward.
We will discuss this approximation scheme further in the context of
expected improvement: § 7.3 p. 127 expected improvement.

6.2 cumulative reward

Simple and global reward are motivated by supposing that the goal of
optimization is to discover the best single point from a space of alterna-
tives. To this end, we evaluate data according to the highest function
value revealed and assume that the values of any suboptimal points
encountered are irrelevant.
In other settings, the value of every individual observation might
be significant, for example, if the optimization procedure is controlling
a critical external system. If the consequences of these decisions are
nontrivial, we might wish to discourage observing where we might
encounter unexpectedly low objective function values.

114
 6.3. information gain

Cumulative reward encourages obtaining observations with large

average value. For a dataset D = (x, y), its cumulative reward is simply
the sum of the observed values:
∑︁
𝑢 (D) = 𝑦𝑖 . (6.7)
𝑖

One notable use of cumulative reward is in active search, a simple active search: § 11.11, p. 282
mathematical model of scientific discovery. Here, we successively select
points for investigation seeking novel members of a rare, valuable class
V ⊂ X. Observing at a point 𝑥 ∈ X yields a binary observation indicating
membership in the desired class: 𝑦 = [𝑥 ∈ V]. Most studies of active
search seek to maximize the cumulative reward (6.7) of the gathered
data, hoping to discover as many valuable items as possible.

6.3 information gain

Simple, global, and cumulative reward judge optimization performance
based solely on having found high objective function values, a natural
and pragmatic concern. Information theory10 provides an alternative 10 A broad introduction to information theory is
approach to measuring utility that is often used in Bayesian optimization. provided by the classical text
An information-theoretic approach to sequential experimental design t. m. cover and j. a. thomas (2006). Elements
(including optimization) identifies some random variable that we wish of Information Theory. John Wiley & Sons,
to learn about through our observations. We then evaluate performance and a treatment focusing on the connections
by quantifying the amount of information about this random variable to Bayesian inference can be found in
revealed by data, favoring datasets containing more information. This
d. j. c. mackay (2003). Information Theory, In-
line of reasoning gives rise to the notion of information gain. ference, and Learning Algorithms. Cambridge
Let 𝜔 be a random variable of interest that we wish to determine University Press.
through the observation of data. The choice of 𝜔 is open-ended and
should be guided by the application at hand. Natural choices aligned
with optimization include the location of the global optimum, 𝑥 ∗, and
the maximal value of the objective, 𝑓 ∗ (1.1), each of which has been
considered in depth in this context.
We may quantify our initial uncertainty about 𝜔 via the (differential) entropy
entropy of its prior distribution, 𝑝 (𝜔):
∫
𝐻 [𝜔] = − 𝑝 (𝜔) log 𝑝 (𝜔) d𝜔.

The information gain offered by a dataset D is then the reduction in 11 A caveat is in order regarding this notation,
entropy when moving from the prior to the posterior distribution: which is not standard. In information theory
𝐻 [𝜔 | D] denotes the conditional entropy of
𝑢 (D) = 𝐻 [𝜔] − 𝐻 [𝜔 | D], (6.8) 𝜔 given D, which is the expectation of the
given quantity over the observed values y. For
our purposes it will be more useful for this
where 𝐻 [𝜔 | D] is the differential entropy of the posterior:11 to signify the differential entropy of the no-
∫ tationally parallel posterior 𝑝 (𝜔 | D). When
𝐻 [𝜔 | D] = − 𝑝 (𝜔 | D) log 𝑝 (𝜔 | D) d𝜔. needed, we will write conditional
 entropy with
an explicit expectation: 𝔼 𝐻 [𝜔 | D] | x .

Somewhat confusingly, some authors use an alternative definition

of information gain – the Kullback–Leibler (kl) divergence between the Kullback–Leibler (kl) divergence

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 115

 utility functions for optimization

posterior distribution and the prior distribution:

∫
12 A simple example: suppose 𝜔 ∈ (0, 1) is the   𝑝 (𝜔 | D)
unknown bias of a coin, with prior 𝑢 (D) = 𝐷 kl 𝑝 (𝜔 | D) k 𝑝 (𝜔) = 𝑝 (𝜔 | D) log d𝜔. (6.9)
𝑝 (𝜔)
𝑝 (𝜔) = Beta(𝜔; 2, 1); 𝐻 ≈ −0.193.
That is, we quantify the information contained in data by how much
After flipping and observing “tails,” the poste- our belief in the 𝜔 changes as a result of collecting it. This definition
rior becomes
has some convenient properties compared to the previous one (6.8);
𝑝 (𝜔 | D) = Beta(𝜔; 2, 2); 𝐻 ≈ −0.125. namely, the expression in (6.9) is is invariant to reparametrization of 𝜔
The information “gained” was and always nonnegative, whereas “surprising” observations may cause
𝐻 [𝜔 ] − 𝐻 [𝜔 | D] ≈ −0.068 < 0.
the information gain in (6.8) to become negative.12 However, the previous
definition as the direct reduction in entropy may be more intuitive.
Of course, the most likely outcome of the flip
Fortunately (and perhaps surprisingly!), there is a strong connection
a priori was “heads,” so the outcome was sur-
prising. Indeed the expected information gain between these two “information gains” (6.8–6.9) in the context of se-
before the experiment was quential decision making. Namely, their expected values with respect to
  observed values are equal, and thus maximizing expected utility with ei-
𝐻 [𝜔 ] − 𝔼 𝐻 [𝜔 | D] ≈ 0.137 > 0.
ther leads to identical decisions.13 For this reason, the reader may simply
13 See p. 138 for a proof. choose whichever definition they find more intuitive.
One-step lookahead with (either) information gain yields an acquisi-
mutual information, entropy search: § 7.6 tion function known as mutual information. This is the basis for a family
p. 135 of related Bayesian optimization procedures sharing the moniker entropy
search, which we will discuss further in the following chapters.
information-theoretic policies as the scientific Unlike the other utility functions discussed thus far, information gain
method: § 7.6, p. 136 is not intimately linked to optimization, and may be adapted to a wide
variety of tasks by selecting the random variable 𝜔 appropriately. Rather,
this scheme of refining knowledge through experiment is effectively a
mathematical formulation of scientific inquiry.

6.4 dependence on model of objective function

One striking feature of most of the utility functions defined in this chapter
is implicit dependence on an underlying model of the objective function.
Both the simple and global reward are defined in terms of the posterior
mean function 𝜇D , and information gain about the location or value of
the optimum is defined in terms of the posterior belief about these values,
𝑝 (𝑥 ,∗ 𝑓 ∗ | D); both of these quantities are byproducts of the objective
function posterior.
model averaging: §§ 4.4–4.5, p. 74 One way to mitigate model dependence in the computation of utility
model-agnostic alternatives is via model averaging (4.11, 4.23).14 We may also attempt to define purely
model-agnostic utility functions in terms of the data alone, without
14 The effect on simple and global reward is to reference to a model; however, the possibilities are somewhat limited if
maximize a model-marginal posterior mean, we wish the resulting utility to be sensible. Cumulative reward (6.2) is
and the effect on information gain is to evalu-
ate changes in model-marginal beliefs about
one example, as it depends only on the observed values y. The maximum
𝜔. function value observed is another possibility (6.6), but, as we have
shown, it is dubious when observations are corrupted by noise. Other
similarly defined alternatives may suffer the same fate – for additive noise
with zero mean, the expected contribution from noise to the cumulative
reward is zero; however, noise will bias many other natural measures
such as order statistics (including the maximum) of the observations.

116
 6.5. comparison of utility functions

samples prior mean prior 95% credible interval

𝑝 (𝑓 ∗ )

𝑝 (𝑥 ∗ )

Figure 6.3: The objective function prior used throughout our utility function comparison. Marginal beliefs of function
values are shown, as well as the induced beliefs over the location of the global optimum, 𝑝 (𝑥 ∗ ), and the value
of the global optimum, 𝑝 (𝑓 ∗ ). Note that there is a significant probability that the global optimum is achieved
on the boundary of the domain, reflected by large point masses.

6.5 comparison of utility functions

We have now presented several utility functions for evaluating a dataset
returned by an optimization routine. Each utility quantifies progress
on our model optimization problem (1.1) in some way, but it may be
difficult at this point to appreciate their, sometimes subtle, differences in
approach. Here we will present and discuss example datasets for which
different utility functions diverge in their opinion of quality.
We particularly wish to contrast the behavior of the simple reward local vs. global properties of posterior
(6.3) with other utility functions. Simple reward is probably the most
prevalent utility in the Bayesian optimization literature (especially in
applications), as it corresponds to the widespread expected improve- expected improvement: § 7.3 p. 127
ment acquisition function. A distinguishing feature of simple reward
is that it evaluates data based only on local properties of the objective
function posterior. This locality is both computationally convenient and
pragmatic. Simple reward is derived from the premise that we will be
recommending one of the points observed during the course of opti-
mization for permanent use, and thus it is sensible to judge performance
based on the objective function values at the observed locations alone.
Several alternatives instead measure global properties of the objective
function posterior. The global reward (6.5), for example, considers the
entire posterior mean function, reflecting a willingless to recommend
an unevaluated point after termination. Information gain (6.8) about
the location or value of the optimum considers the posterior entropy of
these quantities, again a global property. The consequences of reasoning
about local or global properties of the posterior can sometimes lead to
significant disagreement between the simple reward and other utilities.
In the following examples, we consider optimization on an inter- model of objective function
val with exact measurements. We model the objective function with
a Gaussian process with constant mean (3.1) and squared exponential

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 117

 utility functions for optimization

observations posterior mean posterior 95% credible interval

𝑝 (𝑓 ∗ | D)

𝑝 (𝑥 ∗ | D)

Figure 6.4: An example dataset of five observations and the resulting posterior belief of the objective function. This dataset
exhibits relatively low simple reward (6.3) but relatively high global reward (6.5) and information gain (6.8)
about the location 𝑥 ∗ and value 𝑓 ∗ of the optimum.

covariance (3.12). This prior is illustrated in figure 6.3, along with the
induced beliefs about the location 𝑥 ∗ and value 𝑓 ∗ of the global optimum.
15 For this model, a unique optimum will exist Both distributions reflect considerable uncertainty.15 We will examine
with probability one; see § 2.7, p. 34 for more two datasets that might be returned by an optimizer using this model and
details.
discuss how different utility functions would evaluate these outcomes.

Good global outcome but poor local outcome

Consider the dataset in figure 6.4 and the resulting posterior belief about
the objective and its optimum. In this example, the simple reward is
low simple reward relatively low as the posterior mean at our observations is unremarkable.
In fact, every observation was lower than the prior mean, a seemingly
high global reward unlucky outcome. However, the global reward is relatively high: the data
imply a steep derivative in one location, inducing high values of the
posterior mean away from our data. This is a significant accomplishment
from the point of view of the global reward, as the model expects a
terminal recommendation in that region to be especially valuable.
final recommendations Figure 6.1 shows the optimal final recommendations associated with
these two utility functions. The simple reward recommendation prior-
itizes safety over reward, whereas the global reward recommendation
reflects more risk tolerance. Neither is inherently better: although the
global reward recommendation has a larger expected value, this ex-
high information gain pectation is computed using a model that might be mistaken. Further,
comparing the posterior distribution in figure 6.4 with the prior in figure

118
 6.6. summary of major ideas

observations posterior mean posterior 95% credible interval

𝑝 (𝑓 ∗ | D)

𝑝 (𝑥 ∗ | D)

Figure 6.5: An example dataset containing a single observation and the resulting posterior belief of the objective function.
This dataset exhibits relatively high simple reward (6.3) but relatively low global reward (6.5) and information
gain (6.8) about the location 𝑥 ∗ and value 𝑓 ∗ of the optimum.

6.3, we see this example dataset also induces a significant reduction in

our uncertainty about both the location and value of the global optimum,
despite not containing any particularly notable values itself. Therefore,
despite a somewhat low simple reward, observing this dataset results in
relatively high information gain about these quantities.

Good local outcome but poor global outcome

We illustrate a different example dataset in figure 6.5. The dataset contains
a single observation with value somewhat higher than the prior mean.
Although this dataset may not appear particularly impressive, its simple
reward is higher than the previous dataset, as this observation exceeds high simple reward
every value seen in that scenario.
However, this dataset has lower value than the previous dataset when
evaluated other utility functions. Its global reward is lower than in the
first scenario, as the global maximum of the posterior mean is lower. This low global reward
can be verified by visual inspection of figures 6.4 and 6.5, whose vertical
axes are compatible. Further, the single observation in this scenario
provides nearly no information regarding the location nor the value
of the global maximum. The observation of a moderately high value low information gain
provides only weak evidence that the global optimum may be located
nearby, barely influencing our posterior belief about 𝑥 ∗. The observation
does truncate our belief of the value of the global optimum 𝑓,∗ but only
rules out a relatively small portion of its lower tail.

6.6 summary of major ideas

In Bayesian decision theory, preferences over outcomes are encoded by
a utility function, which in the context of optimization policy design,

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 119

 utility functions for optimization

assesses the quality of data returned by an optimization routine, 𝑢 (D).

16 Just like human taste, there is no right or The optimization policy then seeks to design observations to maximize
wrong when it comes to preferences, at least the expected utility of the returned data. The general theory presented
not over certain outcomes. The von Neumann–
Morgenstern theorem mentioned on p. 90 en- in the last chapter makes no assumptions regarding the utility func-
tails rationality axioms, but these only apply tion.16 However, in the context of optimization, some utility functions
to preferences over uncertain outcomes. are particularly easy to motivate.
• In many cases there is a decision following optimization in which we
must recommend a single point in the domain for perpetual use. In this
case, it is sensible to define an optimization utility function in terms of
expected utility of terminal recommendation: the expected utility of the optimal terminal recommendation informed
§ 6.1, p. 109 by the returned data. This requires fully specifying that terminal recom-
mendation, including its action space and utility function, after which
we may “pass through” the optimal expected utility (6.1).
risk tolerance: § 6.1, p. 111 • When designing a terminal recommendation – especially when we may
recommend points with residual uncertainty in their underlying objec-
tive value – it may be prudent to consider our risk tolerance. Careful
design of the terminal utility allows for us to tune our appetite for risk, in
terms of trading off the a point’s expected value against its uncertainty.
Most utilities encountered in Bayesian optimization are risk neutral, but
this need not necessarily be the case.
simple reward: § 6.1, p. 112 • Two notable realizations of this scheme are simple reward (6.3) and global
global reward: § 6.1, p. 113 reward (6.5), both of which represent the expected utility of an optimal
terminal recommendation with a risk-neutral utility. The action space for
simple reward is the points visited during optimization, and the action
space for global reward is the entire domain.
• The simple reward simplifies when observations are exact (6.4).
cumulative reward: § 6.2, p. 114 • An alternative to the simple reward is the cumulative reward (6.7), which
evaluates a dataset based on the average, rather than maximum, value
observed. This does not see too much direct use in policy design, but is
an important concept for the analysis of algorithms.
information gain: § 6.3, p. 115 • Information gain provides an information-theoretic approach to quanti-
fying the value of data in terms of the information provided by the data
regarding some quantity of interest. This can be quantified by either
measuring the reduction in differential entropy moving from the prior
to the posterior (6.8) or by the kl divergence between the posterior and
prior (6.9) – either induces the same one-step lookahead policy.
• In the context of optimization, information gain regarding either the lo-
cation 𝑥 ∗ or value 𝑓 ∗ of the global optimum (1.1) are judicious realizations
of this general approach to utility design.
comparison of utility functions: § 6.5, p. 117 • An important feature distinguishing simple reward from most other
utility functions is its dependence on the posterior belief at the observed
locations alone, rather than the posterior belief over the entire objective
function. Even in relatively simple examples, this may lead to disagree-
ment between simple reward and other utility functions in judging the
quality of a given dataset.

120
 6.6. summary of major ideas

The utility functions presented in this chapter form the backbone

of the most popular Bayesian optimization algorithms. In particular,
many common policies are realized by maximizing the one-step expected one-step lookahead: § 5.3, p. 101
marginal gain to one of these utilities, as we will show in the next chapter.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 121

 7
COMMON BAYESIAN OPTIMIZATION POLICIES

The heart of an optimization routine is its policy, which sequentially

designs each observation in light of available data.1 In the Bayesian
approach to optimization, policies are designed with reference to a prob-
abilistic belief about the objective function, with this belief guiding the
policy in making decisions likely to yield beneficial outcomes. Numer-
ous Bayesian optimization policies have been proposed in the literature,
many of which enjoy widespread use. In this chapter we will present 1 The reader may wish to recall our model opti-
an overview of popular Bayesian optimization policies and emphasize mization procedure: Algorithm 1.1, p. 3.
common themes in their construction. In the next chapter we will pro- computing policies with Gaussian processes:
vide explicit computational details for implementing these policies with chapter 8, p. 157
Gaussian process models of the objective function.
Nearly all Bayesian optimization algorithms result from one of two
primary approaches to policy design. The most popular is Bayesian Bayesian decision theory: chapter 5, p. 87
decision theory, the focus of the previous two chapters. In chapter 5 we
introduced Bayesian decision theory as a general framework for deriving
optimal, but computationally prohibitive, optimization policies. In this
chapter, we will apply the ideas underlying these optimal procedures to
realize computationally tractable and practically useful policies. We will
see that a majority of popular Bayesian optimization algorithms can be one-step lookahead: § 5.3, p. 101
interpreted in a uniform manner as performing one-step lookahead for utility functions for optimization: chapter 6,
some underlying utility function. p. 109
Another avenue for policy design is to adopt algorithms for multi- multi-armed bandits
armed bandits to the optimization setting. A multi-armed bandit is a finite-
dimensional model of sequential optimization with noisy observations.
We consider an agent faced with a finite set of alternatives (“arms”), who
is compelled to select a sequence of items from this set. Choosing a given 2 The name references a gambler contemplating
item yields a stochastic reward drawn from an unknown distribution how to allocate their bankroll among a wall
of slot machines. Slot machines are known as
associated with that arm. We seek a sequential policy for selecting arms “one-armed bandits” in American vernacular,
maximizing the expected cumulative reward (6.2).2 as they eventually steal all your money.
Multi-armed bandits have seen decades of sustained study, and some
policies have strong theoretical guarantees on their performance, sug-
gesting these policies may also be useful for optimization. To this end, we
may model optimization as an infinite-armed bandit, where each point in
the domain 𝑥 ∈ X represents an arm with uncertain reward depending
on the objective function value 𝜙 = 𝑓 (𝑥). Our belief about the objective
function then provides a mechanism to reason about these rewards and 𝜙
derive a policy. This analogy has inspired several Bayesian optimization
policies, many of which enjoy strong performance guarantees. Exploration vs. exploitation. We show reward
A central concern in bandit problems is the exploration–exploitation distributions for two possible options. The
dilemma: we must repeatedly decide whether to allocate resources to an red option returns higher expected reward,
but the blue option reflects more uncertainty
arm already known to yield high reward (“exploitation”) or to an arm with and may actually be superior. Which should
uncertain reward to learn about its reward distribution (“exploration”). we prefer?
Exploitation may yield a high instantaneous reward, but exploration
may provide valuable information for improving future rewards. This 3 p. whittle (1982). Optimization Over Time:
tradeoff between instant payoff and learning for the future has been Dynamic Programming and Stochastic Control.
called “a conflict evident in all human action.”3 A similar choice is faced Vol. 1. John Wiley & Sons.

This material will be published by Cambridge University Press as Bayesian Optimization. 123
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 common bayesian optimization policies

samples observations posterior mean posterior 95% credible interval

Figure 7.1: The scenario we will consider for illustrating optimization policies. The objective function prior is a Gaussian
process with constant mean and Mátern covariance with 𝜈 = 5/2 (3.14). We show the marginal predictive
distributions and three samples from the posterior conditioned on the indicated observations.

throughout optimization, as we must continually decide whether to focus

on a suspected local maximum (exploitation) or to explore unknown
regions of the domain seeking new maxima (exploration). We will see
that typical Bayesian optimization policies reflect consideration of this
dilemma in some way, whether by explicit design on or as a consequence
of decision-theoretic reasoning.
Before diving into policy design, we pause to introduce a running
example we will carry through the chapter and notation to facilitate our
discussion. We will then derive a series of policies stemming from Bayes-
ian decision theory, and finally consider bandit-inspired algorithms.

7.1 example optimization scenario

Throughout this chapter we will demonstrate the behavior of optimiza-
4 Take note of the legend; it will not be repeated. tion policies on an example scenario illustrated in figure 7.1.4 We consider
a one-dimensional objective function observed without noise and adopt
a Gaussian process prior belief about this function. The prior mean
Gaussian processes: chapter 2, p. 15 function is constant (3.1), and the prior covariance function is a Mátern
covariance with 𝜈 = 5/2 (3.14). The parameters are fixed so that the do-
main spans exactly 30 length scales. We condition this prior on three
observations, inducing two local maxima in the posterior mean and a
range of marginal predictive uncertainty.
objective function for simulation We will illustrate the behavior of policies by simulating optimization
to design a sequence of additional observations for this running example.
The ground truth objective function we will use for these simulations is
shown in figure 7.2 and was drawn from the corresponding model. The
objective features numerous undiscovered local maxima and exhibits an
unusually high global maximum on the left-hand side of the domain.

7.2 decision-theoretic policies

utility functions for optimization: chapter 6, Central to decision-theoretic optimization is a utility function 𝑢 (D) mea-
p. 109 suring the quality of a dataset returned by an optimizer. After selecting
a utility function and a model of the objective function and our observa-
optimal policies: § 5.2, p. 91 tions, we may design each observation to maximize the expected utility

124
 7.2. decision-theoretic policies

objective, 𝑓

Figure 7.2: The true objective function used for simulating optimization policies.

of the returned data (5.16). This policy is optimal in the average case: it
maximizes the expected utility of the returned dataset over the space 5 To be precise, optimality is defined with re-
of all possible policies.5 Unfortunately, optimality comes at a great cost. spect to a model for the objective function
𝑝 (𝑓 ), an observation model 𝑝 (𝑦 | 𝑥, 𝜙), a util-
Computing the optimal policy requires recursive simulation of the entire ity function 𝑢 (D), and an upper bound on the
remainder of optimization, a random process due to uncertainty in the number of observations allowed 𝜏. Bayesian
outcomes of our observations. In general, the cost of computing the decision theory provides a policy achieving
optimal policy grows exponentially with the horizon, the number of the maximal expected utility at termination
with respect to these choices.
observations remaining before termination.
However, the structure of the optimal policy suggests a natural family running time of optimal policy and efficient
of lookahead approximations based on fixing a computationally tractable approximations: § 5.3, p. 99
maximum horizon throughout optimization. This line of reasoning has limited lookahead: § 5.3, p. 101
lead to many of the practical policies available for Bayesian optimization.
In fact, most popular algorithms represent one-step lookahead, where
in each iteration we greedily maximize the expected utility after obtain-
ing only a single additional observation. Although these policies are
maximally myopic, they are also maximally efficient among lookahead
approximations and have delivered impressive empirical performance in
a wide range of settings.
It may seem surprising that such dramatically myopic policies have
any use at all. There is a huge difference between the scale of reasoning
in one-step lookahead compared with the optimal procedure, which
may consider hundreds of future decisions or more when designing an
observation. However, the situation is somewhat more nuanced than it
might appear. In a seminal paper, kushner argued that myopic policies
may in fact show better empirical performance than a theoretically
optimal policy, and his argument remains convincing:6

Since a mathematical model of [𝑓 ] is available, it is theoretically 6 h. j. kushner (1964). A New Method of Locat-
possible, once a criterion of optimality is given, to determine the ing the Maximum Point of an Arbitrary Multi-
peak Curve in the Presence of Noise. Journal
mathematically optimal sampling policy. However. . . determination of Basic Engineering 86(1):97–106.
of the optimum sampling policies is extremely difficult. Because
of this, the development of our sampling laws has been guided
primarily by heuristic considerations.7 There are some advantages 7 Specifically, maximizing probability of im-
to the approximate approach. . . [and] its use may yield better results provement: § 7.5, p. 131.
than would a procedure that is optimum for the model. Although the
model selected for [𝑓 ] is the best we have found for our purposes,
it is sometimes too general. . .

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 125

common bayesian optimization policies

What could possibly cause such a seemingly contradictory finding? As

kushner suggests, one possible reason could be model misspecification.
The optimal policy is only defined with respect to a chosen model of the
objective function and our observations, which is bound to be imperfect.
By relying less on the model’s belief, we may gain some robustness
alongside considerable computational savings.
The intimate relationship between many Bayesian optimization meth-
ods and one-step lookahead is often glossed over, with a policy often
introduced ex nihilo and the implied choice of utility function left un-
stated. This disconnect can sometimes lead to policies that are nonsen-
sical from a decision-theoretic perspective or that incorporate implicit
approximations that may not always be appropriate. We intend to clarify
these connections here. We hope that our presentation can help guide
practitioners in navigating the increasingly crowded space of available
policies when presented with a novel scenario.

One-step lookahead
notation for one-step lookahead policies Let us review the generic procedure for developing a one-step lookahead
policy and adopt standard notation to facilitate their description. Suppose
we have selected an arbitrary utility function 𝑢 (D) to evaluate a returned
dataset. Suppose further that we have already gathered an arbitrary
dataset D = (x, y) and wish to select the next evaluation location. This
is the fundamental role of an optimization policy.
proposed next point 𝑥 with putative value 𝑦 If we were to choose some point 𝑥, we would observe a corresponding

updated dataset D0 = D ∪ (𝑥, 𝑦) value 𝑦 and update our dataset, forming D 0 = (x,0 y0) = D ∪ (𝑥, 𝑦) .
Note that in our discussion on decision theory in chapter 5, we notated
this updated dataset with the symbol D1 , as we needed to be able to
distinguish between datasets after the incorporation of a variable number
of additional observations. As our focus in this chapter will be on one-step
lookahead, we can simplify notation by dropping subscripts indicating
time. Instead, we will systematically use the prime symbol to indicate
future quantities after the acquisition of the next observation.
In one-step lookahead, we evaluate a proposed point 𝑥 via the ex-
expected marginal gain pected marginal gain in utility after incorporating an observation there
(5.8):
 
𝛼 (𝑥; D) = 𝔼 𝑢 (D 0) | 𝑥, D − 𝑢 (D),
which serves as an acquisition function inducing preferences over possi-
acquisition functions: § 5, p. 88 ble observation locations. We design each observation by maximizing
this score:
𝑥 ∈ arg max 𝛼 (𝑥 0; D). (7.1)
𝑥 0 ∈X

When the utility function 𝑢 (D) represents the expected utility of

a decision informed by the data, such as a terminal recommendation
following optimization, the expected marginal gain is also known as the
value of sample information value of sample information from observing at 𝑥. This term originates
from the study of decision making in an economic context. Consider

126
 7.3. expected improvement

Table 7.1: Summary of one-step looka-

utility function, 𝑢 (D) expected one-step marginal gain head optimization policies.
simple reward, (6.3) expected improvement, § 7.3
global reward, (6.5) knowledge gradient, § 7.4
improvement to simple reward probability of improvement, § 7.5
information gain, (6.8) or (6.9) mutual information, § 7.6
cumulative reward, (6.7) posterior mean, § 7.10

an agent who must make a decision under uncertainty, and suppose

they have access to a third party who is willing to provide potentially
insightful advice in exchange for a fee. By reasoning about the potential
impact of this advice on the ultimate decision, we may quantify the 8 j. marschak and r. radner (1972). Economic
expected value of the information,8, 9 and determine whether the offered Theory of Teams. Yale University Press. [§ 2.12]
advice is worth the investment. 9 h. raiffa and r. schlaifer (1961). Applied Sta-
Due to its simplicity and inherent computational efficiency, one-step tistical Decision Theory. Division of Research,
Graduate School of Business Administration,
lookahead is a pervasive approximation scheme in Bayesian optimiza- Harvard University. [§ 4.5]
tion. Table 7.1 provides a list of common acquisition functions, each
representing the expected one-step marginal gain to a corresponding
utility function. We will discuss each in detail below.

7.3 expected improvement

Adopting the simple reward utility function (6.3) and performing one- simple reward: § 6.1, p. 109
step lookahead defines the expected improvement acquisition function.
Sequential maximization of expected improvement is perhaps the most
widespread policy in all of Bayesian optimization.
Suppose that we wish to locate a single location in the domain with
the highest possible objective value and ultimately wish to recommend
one of the points investigated during optimization for permanent use.
The simple reward utility function evaluates a dataset D precisely by
the expected value of an optimal final recommendation informed by the
data, assuming risk neutrality: risk neutrality: § 6.1, p. 109

𝑢 (D) = max 𝜇D (x). 10 This reasoning is the same for all one-step
lookahead policies, which could all be de-
Suppose we have already gathered observations D = (x, y) and wish scribed as maximizing “expected improve-
to choose the next evaluation location. Expected improvement is derived ment.” But this name has been claimed for the
by measuring the expected marginal gain in utility, or the instantaneous simple reward utility alone.

improvement, 𝑢 (D 0) − 𝑢 (D),10 offered by making the next observation 11 As mentioned in the last chapter, simple re-
ward degenerates with an empty dataset; ex-
at a proposed location 𝑥:11
pected improvement does as well. In that case
∫
  we can simply ignore the second term and
𝛼 ei (𝑥; D) = max 𝜇D0 (x0) 𝑝 (𝑦 | 𝑥, D) d𝑦 − max 𝜇D (x). (7.2) compute the first, which for zero-mean addi-
tive noise becomes the mean function of the
prior process.
Expected improvement reduces to a particularly nice expression in
the case of exact observations of the objective, where the utility takes a
simpler form (6.4). Suppose that, when we elect to make an observation expected improvement without noise
at a location 𝑥, we observe the exact objective value 𝜙 = 𝑓 (𝑥). Consider

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 127

 common bayesian optimization policies

expected improvement, 𝛼 ei next observation location

Figure 7.3: The expected improvement acquisition function (7.2) corresponding to our running example.

maximal value observed, incumbent 𝜙 ∗ a dataset D = (x, 𝝓), and define 𝜙 ∗ = max 𝝓 to be the so-called incum-
bent: the maximal objective value yet seen.12 As a consequence of exact
12 The value 𝜙 ∗ is incumbent as it is currently observation, we have
“holding office” as our standing recommenda-
tion until it is deposed by a better candidate. 𝑢 (D) = 𝜙 ∗ ; 𝑢 (D 0) = max(𝜙 ,∗ 𝜙);
and thus
𝑢 (D 0) − 𝑢 (D) = max(𝜙 − 𝜙 ,∗ 0).

Substituting into (7.2), in the noiseless case we have

∫
𝛼 ei (𝑥; D) = max(𝜙 − 𝜙 ,∗ 0) 𝑝 (𝜙 | 𝑥, D) d𝜙 . (7.3)

example and interpretation Expected improvement is illustrated for our running example in
figure 7.3. In this case, maximizing expected improvement will select
a point near the previous best point found, an example of exploitation.
Notice that the expected improvement vanishes near regions where
we have existing observations. Although these locations may be likely
to yield values higher than 𝜙 ∗ due to relatively high expected value,
the relatively narrow credible intervals suggest that the magnitude of
any improvement is likely to be small. Expected improvement is thus
considering the exploration–exploitation dilemma in the selection of the
next observation location, and the tradeoff between these two concerns
is considered automatically.
simulated optimization and interpretation Figure 7.4 shows the posterior belief of the objective after sequentially
maximizing expected improvement to gather 20 additional observations
of our example objective function. The global optimum was efficiently
located. The distribution of the sample locations, with more evaluations
in the most promising regions, reflects consideration of the exploration–
exploitation dilemma. However, there seems to have been a focus on
exploitative behavior resulting from myopia exploitation throughout the entire process; the first ten observations
for example never strayed from the initially known local optimum. This
behavior is a reflection of the simple reward utility function underlying
the policy, which only rewards the discovery of high objective func-
tion values at observed locations. As a result, one-step lookahead may

128
 7.4. knowledge gradient

Figure 7.4: The posterior after 10 (top) and 20 (bottom) steps of the optimization policy induced by the expected improve-
ment acquisition function (7.2) on our running example. The tick marks show the points chosen by the policy,
progressing from top to bottom, during iterations 1–10 (top) and 11–20 (bottom). Observations within 0.2
length scales of the optimum are marked in red; the optimum was located on iteration 19.

rationally choose to make marginal improvements to the value of the

best-seen point, even if the underlying function value is known with a
fair amount of confidence.

7.4 knowledge gradient

Adopting the global reward utility (6.5) and performing one-step-lookahead global reward: § 6.1, p. 109
yields an acquisition function known as the knowledge gradient.
Assume that, just as in the situation leading to the derivation of
expected improvement, we again wish to identify a single point in the
domain maximizing the objective function. However, imagine that at
termination we are willing to commit to a location possibly never evalu-
ated during optimization. To this end, we adopt the global reward utility
function to measure our progress:
𝑢 (D) = max 𝜇D (𝑥),
𝑥 ∈X

which rewards data for increasing the posterior mean, irrespective of

location. Computing the one-step marginal gain to this utility results in
the knowledge gradient acquisition function:
∫ 
𝛼 kg (𝑥; D) = max
0
𝜇D0 (𝑥 ) 𝑝 (𝑦 | 𝑥, D) d𝑦 − max
0
0
𝜇D (𝑥 0). (7.4)
𝑥 ∈X 𝑥 ∈X
13 p. frazier and w. powell (2007). The Knowl-
The knowledge gradient moniker was coined by frazier and pow- edge Gradient Policy for Offline Learning with
Independent Normal Rewards. adprl 2007.
ell,13 who interpreted the global reward as the amount of “knowledge”

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 129

common bayesian optimization policies

knowledge gradient, 𝛼 kg next observation location

Figure 7.5: The knowledge gradient acquisition function (7.4) corresponding to our running example.

samples of updated posterior mean, 𝜇 D0

Figure 7.6: Samples of the updated posterior mean
when evaluating at the location chosen
by the knowledge gradient, illustrated in
figure 7.5. Only the right-hand section of
the domain is shown.

about the global maximum offered by a dataset D. The knowledge gradi-

ent 𝛼 kg (𝑥; D) can then be interpreted as the expected change in knowl-
edge (that is, a discrete-time gradient) offered by a measurement at 𝑥.
example and interpretation The knowledge gradient is illustrated for our running example in
figure 7.5. Perhaps surprisingly, the chosen observation location is re-
markably close to the previously best-seen point. At first glance, this
may seem wasteful, as we are already fairly confident about the value
we might observe.
reason for selected observation However, the knowledge gradient seeks to maximize the global max-
imum of the posterior mean, regardless of its location. With this in mind,
we may reason as follows. There must be a local maximum of the objec-
tive function in the neighborhood of the best-seen point, but our current
knowledge is insufficient to pinpoint its location. Further, as the rele-
vant local maximum is probably not located precisely at this point, the
objective function is either increasing or decreasing as it passes through.
If we were to learn the derivative of the objective at this point, we would
adjust our posterior belief to reflect that knowledge. Regardless of the
sign or exact value of the derivative, our updated belief would reflect
the discovery of a new, higher local maximum of the posterior mean
in the indicated direction. By evaluating at the location selected by the
knowledge gradient, we can effectively estimate the derivative of the
objective; this is the principle behind finite differencing.
In figure 7.6, we show samples of the updated posterior mean func-
tion 𝜇D0 (𝑥) derived from sampling from the predictive distribution at
the chosen evaluation location and conditioning. Indeed, these samples
exhibit newly located global maxima on either side of the selected point,
depending on the sign of the implied derivative. Note that the locations

130
 7.5. probability of improvement

Figure 7.7: The posterior after 10 (top) and 20 (bottom) steps of the optimization policy induced by the knowledge gradient
acquisition function (7.4) on our running example. The tick marks show the points chosen by the policy,
progressing from top to bottom, during iterations 1–10 (top) and 11–20 (bottom). Observations within 0.2 length
scales of the optimum are marked in red; the optimum was located on iteration 15.

of these new maxima coincide with local maxima of the expected im-
provement acquisition function; see figure 7.3 for comparison. This is
not a coincidence! One way to interpret this relation is that, due to re-
warding large values of the posterior mean at observed locations only,
expected improvement must essentially guess on which side the hidden
local optimum of the objective lies and hope to be correct. The knowl-
edge gradient, on the other hand, considers identifying this maximum
on either side a success, and guessing is not necessary.
Figure 7.7 illustrates the behavior of the knowledge gradient policy simulated optimization and interpretation
on our example optimization scenario. The global optimum was located
efficiently. Comparing the decisions made by the knowledge gradient to
those made by expected improvement (see figure 7.4), we can observe more exploration than expected improvement
a somewhat more even exploration of the domain, including in local from more-relaxed utility
maxima. The knowledge gradient policy does not necessarily need to
expend observations to verify a suspected maximum, instead putting
more trust into the model to have correct beliefs in these regions.

7.5 probability of improvement

As its name suggests, the probability of improvement acquisition function
computes the probability of an observed value to improve upon some
chosen threshold, regardless of the magnitude of this improvement.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 131

common bayesian optimization policies

improvement target
𝑝 (𝜙 | 𝑥, D)
𝑝 (𝜙 0 | 𝑥 ,0 D)

Figure 7.8: An illustrative example comparing the behavior of probability of improvement with expected improvement
computed with respect to the dashed target. The predictive distributions for two points 𝑥 and 𝑥 0 are shown. The
distributions have equal mean but the distribution at 𝑥 0 has larger predictive standard deviation. The shaded
regions represent the region of improvement. The relatively safe 𝑥 is preferred by probability of improvement,
whereas the more-risky 𝑥 0 is preferred by expected improvement.

simple reward: § 6.1, p. 109 Consider the simple reward of an already gathered dataset D = (x, y):

𝑢 (D) = max 𝜇D (x).

The probability of improvement acquisition function scores a proposed

observation location 𝑥 according to the probability that an observation
desired margin of improvement, 𝜀 there will improve this utility by at least some margin 𝜀. Let us denote the
desired improvement threshold, 𝜏 desired utility threshold with 𝜏 = 𝑢 (D) + 𝜀; we will use both the absolute
threshold 𝜏 and the marginal threshold 𝜀 in the following discussion as
convenient. The probability of improvement is then the probability that
the updated utility 𝑢 (D 0) exceeds the chosen threshold:


𝛼 pi (𝑥; D, 𝜏) = Pr 𝑢 (D 0) > 𝜏 | 𝑥, D . (7.5)

utility formulation We may interpret probability of improvement in the Bayesian decision-

theoretic framework as computing the expected one-step marginal gain
in a peculiar choice of utility function: a utility offering unit reward for
each observation increasing the simple reward by the desired amount.
noiseless case In the case of exact observation, we have

𝑢 (D) = max 𝑓 (x) = 𝜙 ∗; 𝑢 (D 0) = max(𝜙 ,∗ 𝜙),

and we may write the probability of improvement in the somewhat

simpler form
𝛼 pi (𝑥; D, 𝜏) = Pr(𝜙 > 𝜏 | 𝑥, D). (7.6)
In this case, the probability of improvement is simply the complementary
cumulative distribution function of the predictive distribution evaluated
at the improvement threshold 𝜏. This form of probability of improvement
is sometimes encountered in the literature, but our modification in terms
of the simple reward allows for inexact observations as well.
comparison with expected improvement It can be illustrative to compare the preferences over observation
locations implied by the probability of improvement and expected im-
provement acquisition functions. In general, probability of improvement
is somewhat more risk-averse than expected improvement, because prob-
ability of improvement would prefer a certain improvement of modest

132
 7.5. probability of improvement

probability of improvement, 𝛼 pi (𝜀 = 0) next observation location

𝛼 pi (𝜀 = 0.1)

𝛼 pi (𝜀 = 1)

Figure 7.9: The probability of improvement acquisition function (7.5) corresponding to our running example for different
values of the target improvement 𝜀. The target is expressed as a fraction of the range of the posterior mean
over the space. Increasing the target improvement leads to increasingly exploratory behavior.

magnitude to an uncertain improvement of potentially large magnitude.

Figure 7.8 illustrates this phenomenon. Shown are the predictive distri-
butions for the objective function values at two points 𝑥 and 𝑥 .0 Both
points have equal predictive means; however, 𝑥 0 has a significantly larger
predictive standard deviation. We consider improvement with respect
to the illustrated target. The shaded regions represent the regions of
improvement; the probability mass of these regions equal the probabil-
ities of improvement. Improvement is near certain at 𝑥 (𝛼 pi = 99.9%),
whereas it is somewhat smaller at 𝑥 0 (𝛼 pi = 72.6%), and thus probability
of improvement would prefer to observe at 𝑥. The expected improvement
at 𝑥, however, is small compared to 𝑥 0 with its longer tail:
𝛼 ei (𝑥 0; D)
= 1.28.
𝛼 ei (𝑥; D)
The expected improvement at 𝑥 0 is 28% larger than at 𝑥, indicating a
preference for a less-certain but potentially larger payout.

The role of the improvement target

The magnitude of the required improvement plays a crucial role in shap-
ing the behavior of probability of improvement policies. By adjusting
this parameter, we may encourage exploration (with large 𝜀) or exploita-
tion (with small 𝜀). Figure 7.9 shows the probability of improvement for

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 133

common bayesian optimization policies

Figure 7.10: The posterior after 10 (top)

 and 20 (bottom) steps of the optimization policy induced by probability of
improvement with 𝜀 = 0.1 max 𝜇 D (𝑥) − min 𝜇 D (𝑥) (7.5) on our running example. The tick marks show the
points chosen by the policy, progressing from top to bottom, during iterations 1–10 (top) and 11–20 (bottom).
Observations within 0.2 length scales of the optimum are marked in red; the optimum was located on iteration
15.

our example scenario with thresholds corresponding to infinitesimal im-

provement, a modest improvement, and a significant improvement. The
shift towards exploratory behavior for larger improvement thresholds
can be clearly seen.
In figure 7.10, we see 20 evaluations chosen by maximizing probability
of improvement with the target dynamically set to 10% of the range
of the posterior mean function. The global optimum was located, and
the domain appears sufficiently explored. Although performance was
quite reasonable here, the improvement threshold was set somewhat
arbitrarily, and it is not always clear how one should set this parameter.
the 𝜀 = 0 case On one extreme, some authors define a parameter-free (and prob-
ably too literal) version of probability of improvement by fixing the
improvement target to 𝜀 = 0, rewarding even infinitesimal improvement
to the current data. Intuitively, this low bar can induce overly exploita-
tive behavior. Examining the probability of improvement with 𝜀 = 0
for our running example in figure 7.9, we see that the acquisition func-
tion is maximized directly next to the previously best-found point. This
decision represents extreme exploitation and potentially undesirable
behavior. The situation after applying probability of improvement with
𝜀 = 0 to select 20 additional observation locations, shown in figure 7.11,
clearly demonstrates a drastic focus on exploitation. Notably, the global
optimum was not identified.

134
 7.6. mutual information and entropy search

Figure 7.11: The posterior after 20 steps of the optimization policy induced by probability of improvement with 𝜀 = 0
(7.5) on our running example. The tick marks show the points chosen by the policy, progressing from top to
bottom.

Evidently we must carefully select the desired improvement threshold 14 d. r. jones (2001). A Taxonomy of Global Op-
to achieve ideal behavior. jones provided some simple, data-driven advice timization Methods Based on Response Sur-
faces. Journal of Global Optimization 21(4):345–
for choosing improvement thresholds that remains sound.14 Define 383.

𝜇 ∗ = max 𝜇D (𝑥); 𝑟 = max 𝜇D (x) − min 𝜇D (x);

𝑥 ∈X

to represent the global maximum of the posterior mean and the range of
the posterior mean at the observed locations. jones suggests considering
targets of the form
𝜇 ∗ + 𝛼𝑟,
where 𝛼 ≥ 0 controls the amount of desired improvement in terms of 15 The proposed values for 𝛼 given by jones are
the range of observed data. He provides a table of 27 suggested values compiled below.
for 𝛼 in the range [0, 3] and remarks that the points optimizing the set 𝛼
of induced acquisition functions typically cluster together in a small 0 0.07 0.25
number of locations, each representing a different tradeoff between 0.0001 0.08 0.3
0.001 0.09 0.4
exploration and exploitation.15 jones continued to recommend selecting 0.01 0.1 0.5
one point from each of these clusters to evaluate in parallel, defining a 0.02 0.11 0.75
batch optimization policy. Although this may not always be possible, the 0.03 0.12 1
recommended parameterization of the desired improvement is natural 0.04 0.13 1.5
0.05 0.15 2
and would be appropriate for general use. 0.06 0.2 3
This proposal is illustrated for our running example in figure 7.12. We
begin with the posterior after selecting 10 points in our previous demo
(see figure 7.10), and indicate the points maximizing the probability of
improvement for jones’s proposed improvement targets. The points clus-
ter together in four regions reflecting varying exploration–exploitation
tradeoffs.

7.6 mutual information and entropy search

A family of information-theoretic optimization policies have been pro-
posed in recent years, most with variations on the name entropy search.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 135

 common bayesian optimization policies

Figure 7.12: The points maximizing probability of improvement using the 27 improvement thresholds proposed by
jones, beginning with the posterior from figure 7.10 after 10 total observations have been obtained. The tick
marks show the chosen points and cluster together in four regions representing different tradeoffs between
exploration and exploitation.

16 t. m. cover and j. a. thomas (2006). Elements The acquisition function in these methods is mutual information, a mea-
of Information Theory. John Wiley & Sons. sure of dependence between random variables that is a central concept
17 d. j. c. mackay (2003). Information Theory, In- in information theory.16, 17
ference, and Learning Algorithms. Cambridge The reasoning underlying entropy search policies is somewhat dif-
University Press.
ferent from and more general than the other acquisition functions we
have considered thus far, all of which ultimately focus on maximizing
the posterior mean function. Although this is a pragmatic concern, it is
information-theoretic decision making as a intimately linked to optimization. Information-theoretic experimental
model of the scientific method design is instead motivated by an abstract pursuit of knowledge, and may
be interpreted as a mathematical formulation of the scientific method.
We begin by identifying some unknown feature of the world that
we wish to learn about; in the context of Bayesian inference, this will be
some random variable 𝜔. We then view each observation we make as an
opportunity to learn about this random variable, and seek to gather data
that will, in aggregate, provide considerable information about 𝜔. This
process is analogous to a scientist designing a sequence of experiments
to understand some natural phenomenon, where each experiment may
be chosen to challenge or confirm constantly evolving beliefs.
The framework of information theory allows us to formalize this
process. We may quantify the amount of information provided about a
information gain: § 6.3, p. 115 random variable 𝜔 by a dataset D via the information gain, a concept for
which we provided two definitions in the last chapter. Adopting either
definition as a utility function and performing one-step lookahead yields
mutual information as an acquisition function.
Information-theoretic optimization policies select 𝜔 such that its de-
termination gives insight into our optimization problem (1.1). However,
by selecting different choices for 𝜔, we can generate radically differ-
ent policies, each attempting to learn about a different aspect of the
system of interest. Maximizing mutual information has long been pro-
18 d. v. lindley (1956). On a Measure of the Infor- moted as a general framework for optimal experimental design,18 and
mation Provided by an Experiment. The An- this framework has been applied in numerous active learning settings.19
nals of Mathematical Statistics 27(4):986–1005.
Before showing how mutual information arises in a decision-theoretic
19 b. settles (2012). Active Learning. Morgan & context, we pause to define the concept and derive some important prop-
Claypool.
erties.

136
 7.6. mutual information and entropy search

Mutual information
Let 𝜔 and 𝜓 be random variables with probability density functions 𝑝 (𝜔) definition
and 𝑝 (𝜓 ). The mutual information between 𝜔 and 𝜓 is
∬
𝑝 (𝜔,𝜓 )
𝐼 (𝜔;𝜓 ) = 𝑝 (𝜔,𝜓 ) log d𝜔 d𝜓 . (7.7)
𝑝 (𝜔) 𝑝 (𝜓 )
This expression may be recognized as the Kullback–Leibler divergence
between the joint distribution of the random variables and the product
of their marginal distributions:
 
𝐼 (𝜔;𝜓 ) = 𝐷 kl 𝑝 (𝜔,𝜓 ) k 𝑝 (𝜔) 𝑝 (𝜓 ) .

We may extend this definition to conditional probability distributions as conditional mutual information, 𝐼 (𝜔;𝜓 | D)
well. Given an arbitrary set of observed data D, we define the conditional
mutual information between 𝜔 and 𝜓 by:20 20 Some authors use the notation 𝐼 (𝜔;𝜓 | D) to
∬ represent the expectation of the given quantity
𝑝 (𝜔,𝜓 | D) with respect to the dataset D. In optimization,
𝐼 (𝜔;𝜓 | D) = 𝑝 (𝜔,𝜓 | D) log d𝜔 d𝜓 . we will always have an explicit dataset in hand,
𝑝 (𝜔 | D) 𝑝 (𝜓 | D)
in which case the provided definition is more
useful.
Here we have simply conditioned all distributions on the data and applied
the definition in (7.7) to the posterior beliefs.
Several properties of mutual information are immediately evident symmetry
from its definition. First, mutual information is symmetric in its argu-
ments:
𝐼 (𝜔;𝜓 ) = 𝐼 (𝜓 ; 𝜔). (7.8)
We also have that if 𝜔 and 𝜓 are independent, then 𝑝 (𝜔,𝜓 ) = 𝑝 (𝜔) 𝑝 (𝜓 )
and the mutual information is zero:
∬
𝑝 (𝜔) 𝑝 (𝜓 )
𝐼 (𝜔;𝜓 ) = 𝑝 (𝜔,𝜓 ) log d𝜔 d𝜓 = 0.
𝑝 (𝜔) 𝑝 (𝜓 )

Further, recognition of mutual information as a Kullback–Leibler diver- nonnegativity

gence implies several additional inherited properties, including nonneg-
ativity. Thus mutual information attains its minimal value when 𝜔 and
𝜓 are independent.
We may also manipulate (7.7) by twice applying the identity expected reduction in entropy

𝑝 (𝜔,𝜓 ) = 𝑝 (𝜓 ) 𝑝 (𝜔 | 𝜓 )

to derive an equivalent expression for the mutual information:

∬
𝑝 (𝜔,𝜓 )
𝐼 (𝜔;𝜓 ) = 𝑝 (𝜔,𝜓 ) log d𝜔 d𝜓
𝑝 (𝜔) 𝑝 (𝜓 )
∬ ∫
= 𝑝 (𝜔,𝜓 ) log 𝑝 (𝜔 | 𝜓 ) d𝜔 d𝜓 − 𝑝 (𝜔) log 𝑝 (𝜔) d𝜔
∫ ∫ 
= 𝑝 (𝜓 ) 𝑝 (𝜔 | 𝜓 ) log 𝑝 (𝜔 | 𝜓 ) d𝜔 d𝜓 + 𝐻 [𝜔]
 
= 𝐻 [𝜔] − 𝔼 𝐻 [𝜔 | 𝜓 ] . (7.9)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 137

 common bayesian optimization policies

21 It is important to note that this is true only in Thus the mutual information between 𝜔 and 𝜓 is the expected decrease
expectation. Consider two random variables 𝑥 in the differential entropy of 𝜔 if we were to observe 𝜓 . Due to symmetry
and 𝑦 with the following joint distribution. 𝑥
takes value 0 or 1 with probability 1/2 each. If
(7.8), we may swap the roles of 𝜔 and 𝜓 to derive an equivalent expression
𝑥 is 0, 𝑦 takes value 0 or 1 with probability 1/2 in the other direction:
each. If 𝑥 is 1, 𝑦 takes value 0 or −1 with proba-    
bility 1/2 each. The entropy of 𝑥 is 1 bit and the 𝐼 (𝜔;𝜓 ) = 𝐻 [𝜔] − 𝔼𝜓 𝐻 [𝜔 | 𝜓 ] = 𝐻 [𝜓 ] − 𝔼𝜔 𝐻 [𝜓 | 𝜔] . (7.10)
entropy of 𝑦 is 1.5 bits. Observing 𝑥 always
yields 0.5 bits about 𝑦. However, observing Observing either 𝜔 or 𝜓 will, in expectation, provide the same amount
𝑦 produces either no information about 𝑥 (0 of information about the other: the mutual information 𝐼 (𝜔;𝜓 ).21
bits), with probability 1/2, or complete infor-
mation about 𝑥 (1 bit), with probability 1/2.
So the information gain about 𝑥 from 𝑦 and Maximizing mutual information as an optimization policy
about 𝑦 from 𝑥 is actually never equal. How-
ever, the expected information gain is equal, Mutual information arises naturally in Bayesian sequential experimental
𝐼 (𝑥; 𝑦) = 0.5 bits. design as the one-step expected information gain resulting from an ob-
servation. In the previous chapter, we introduced two different methods
for quantifying this information gain. The first was the reduction in the
differential entropy of 𝜔 from the prior to the posterior:
𝑢 (D) = 𝐻 [𝜔] − 𝐻 [𝜔 | D] (7.11)
∫ ∫
= 𝑝 (𝜔 | D) log 𝑝 (𝜔 | D) d𝜔 − 𝑝 (𝜔) log 𝑝 (𝜔) d𝜔.

The second was the Kullback–Leibler divergence between the posterior

  and the prior:
22 Setting 𝑢 (D) = 𝐷 kl 𝑝 (𝜔 | D) k 𝑝 (𝜔) , we
∫
have:   𝑝 (𝜔 | D)
  𝑢 (D) = 𝐷 kl 𝑝 (𝜔 | D) k 𝑝 (𝜔) = 𝑝 (𝜔 | D) log d𝜔. (7.12)
𝔼 𝑢 (D0 ) | 𝑥, D 𝑝 (𝜔)
∫ 

= 𝔼 𝑝 (𝜔 | D0 ) log 𝑝 (𝜔 | D0 ) d𝜔  𝑥, D Remarkably, performing one-step lookahead with either choice yields
∫ mutual information as an acquisition function.
− 𝑝 (𝜔 | D) log 𝑝 (𝜔) d𝜔 Let us first compute the expected marginal gain in (7.11). In this case
  the marginal information gain is:
= −𝔼 𝐻 [𝜔 | 𝑥, D0 ] | D
∫
− 𝑝 (𝜔 | D) log 𝑝 (𝜔) d𝜔.
𝐻 [𝜔 | D] − 𝐻 [𝜔 | D 0],

Here the second term is known as the cross

and the expected marginal information gain is then:
entropy between 𝑝 (𝜔) and 𝑝 (𝜔 | D). We can  
𝛼 mi (𝑥; D) = 𝐻 [𝜔 | D] − 𝔼 𝐻 [𝜔 | D 0] | 𝑥, D (7.13)
also rewrite the utility in similar terms:
∫ = 𝐼 (𝑦; 𝜔 | 𝑥, D),
log 𝑝 (𝜔 | D)
𝑢 (D) = 𝑝 (𝜔 | D) d𝜔
log 𝑝 (𝜔) where we have recognized the expected reduction in entropy in (7.13)
∫
= 𝑝 (𝜔 | D) log 𝑝 (𝜔 | D) d𝜔 as the mutual information between 𝑦 and 𝜔 given the putative location
∫ 𝑥 and the available data D (7.9). It is simple to verify that the expected
− 𝑝 (𝜔 | D) log 𝑝 (𝜔) d𝜔 marginal improvement to the alternative information gain definition
= −𝐻 [𝜔 | D] (7.12) gives the same expression; several terms cancel when computing
∫ the expectation, and those that remain are identical to those in (7.13).22
− 𝑝 (𝜔 | D) log 𝑝 (𝜔) d𝜔. Due to the symmetry of mutual information, we have several equiv-
If we subtract, the cross-entropy terms cancel alent forms for this acquisition function (7.10):
and we obtain mutual information:
𝛼 mi (𝑥; D) = 𝐼 (𝑦; 𝜔 | 𝑥, D)
   
𝔼 𝑢 (D0 ) | 𝑥, D − 𝑢 (D) = = 𝐻 [𝜔 | D] − 𝔼𝑦 𝐻 [𝜔 | D 0] | 𝑥, D (7.14)
   
𝐻 [𝜔 | D] − 𝔼 𝐻 [𝜔 | D0 ] | D .
= 𝐻 [𝑦 | 𝑥, D] − 𝔼𝜔 𝐻 [𝑦 | 𝜔, 𝑥, D] | 𝑥, D . (7.15)

138
 7.6. mutual information and entropy search

𝑝 (𝑓 ∗ | D)

𝑝 (𝑥 ∗ | D)

Figure 7.13: The posterior belief about the location of the global optimum, 𝑝 (𝑥 ∗ | D), and about the value of the global
optimum, 𝑝 (𝑓 ∗ | D), for our running example. Note the significant probability mass associated with the
optimum lying on the boundary.

Depending on the application, one of these two forms may be preferable,

and maximizing either results in the same policy.
Adopting mutual information as an acquisition function for optimiza-
tion requires that 𝜔 be selected to support the optimization task. Two
natural options present themselves: the location of the global optimum,
𝑥 ∗, and the maximum value attained, 𝑓 ∗ = 𝑓 (𝑥 ∗ ) (1.1). Both have received
extensive consideration, and we will discuss each in turn.

Mutual information with 𝑥 ∗

Several authors have proposed mutual information with the location of 23 We tacitly assume the location of the global op-
the global optimum 𝑥 ∗ as an acquisition function:23 timum is unique to simplify discussion. Tech-
nically 𝑥 ∗ is a set-valued random variable. For
Gaussian process models, the uniqueness of
𝛼𝑥 ∗ (𝑥; D) = 𝐼 (𝑦; 𝑥 ∗ | 𝑥, D). (7.16) 𝑥 ∗ can be guaranteed under mild assumptions
(§ 2.7, p. 34).
The distribution of 𝑥 ∗ is illustrated for our running example in figure
7.13. Even for this simple example, the distribution of the global optimum
is nontrivial and multimodal. In fact, in this case, there is a significant
probability that the global maximum occurs on the boundary of the 24 A proper treatment would separate the prob-
domain, which has Lebesgue measure zero, so 𝑥 ∗ does not even have ability density on the interior of X from the
distribution restricted to the boundary, but in
a proper probability density function.24 We will nonetheless use the practice we will only ever be sampling from
notation 𝑝 (𝑥 ∗ | D) in our discussion below. this distribution, as it is simply too compli-
The middle panel of figure 7.14 shows the mutual information with cated to work with directly.
𝑥 ∗ (7.16) for our running example. The next evaluation location will
be chosen to investigate the neighborhood of the best-seen point. It is
interesting to compare the behavior of the mutual information with other
acquisition functions near the boundary of the domain. Although there is
a significant probability that the maximum is achieved on the boundary,
mutual information indicates that we cannot expect to reveal much
information regarding 𝑥 ∗ by measuring there. More information tends to
be revealed by evaluating away from the boundary, as we can reduce our

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 139

 common bayesian optimization policies

mutual information with 𝑥 ∗, 𝛼𝑥 ∗ next observation location

mutual information with 𝑓,∗ 𝛼 𝑓 ∗ next observation location

Figure 7.14: The mutual information between the observed value and the location of the global optimum, 𝛼𝑥 ∗ (middle
panel), and between the observed value and the value of the global optimum, 𝛼 𝑓 ∗ (bottom panel), for our
running example.

uncertainty about the objective function over a larger volume. Expected

improvement (figure 7.3) and probability of improvement (figure 7.9),
on the other hand, are computed only from inspection of the marginal
predictive distribution 𝑝 (𝑦 | 𝑥, D). As a result, they cannot differentiate
observation locations based on their global impact on our belief.
In figure 7.15, we demonstrate 20 steps of optimization by maximizing
the mutual information with 𝑥 ∗ (7.16) for our scenario. We also show the
posterior belief about the maximum location at termination, 𝑝 (𝑥 ∗ | D).
The global optimum was discovered efficiently and with remarkable con-
fidence. Further, the posterior mode matches the true optimal location.

Mutual information with 𝑓 ∗

25 Again we assume in this discussion that a Mutual information with the value of the global optimum 𝑓 ∗ has also
global optimum 𝑓 ∗ exists almost surely. This been investigated as an acquisition function:25
assumption can be guaranteed for Gaussian
process models under mild assumptions (§ 2.7,
𝛼 𝑓 ∗ (𝑥; D) = 𝐼 (𝑦; 𝑓 ∗ | 𝑥, D). (7.17)
p.34).
The distribution of this quantity is illustrated for our running example
in figure 7.13. There is a sharp mode in the distribution corresponding
to the best-seen value in fact being near-optimal, and there is no mass
below the best-seen value, as it serves as a lower bound on the maximum
due to the assumption of exact observation.
example and discussion The bottom panel of figure 7.14 shows the mutual information with
𝑓 ∗ (7.17) for our running example. The next evaluation location will be
chosen to investigate the neighborhood of the best-seen point. It is inter-
esting to contrast this surface with that of the mutual information with
𝑥 ∗ in figure (7.16). Mutual information with 𝑓 ∗ is heavily penalized near

140
 7.7. multi-armed bandits and optimization

𝑝 (𝑥 ∗ | D)

Figure 7.15: The posterior after 10 (top) and 20 (bottom) steps of maximizing the mutual information between the observed
value 𝑦 and the location of the global maximum 𝑥 ∗ (7.16) on our running example. The tick marks show the
points chosen by the policy, progressing from top to bottom. Observations within 0.2 length scales of the
optimum are marked in red; the optimum was located on iteration 16.

existing observations, even those with relatively high values. Observing

at these points would contribute relatively little information about the
value of the optimum, as their predictive distributions already reflect
a narrow range of possibilities and contribute little to the distribution
of 𝑓 .∗ Further, points on the boundary were less favored when seeking
to learn about 𝑥 ∗. However, these points are expected to provide just as
much information about 𝑓 ∗ as neighboring points.
Figure 7.16 illustrates 25 evaluations chosen by sequentially maximiz-
ing the mutual information with 𝑓 ∗ (7.17) for our scenario, along with
the posterior belief about the value of the maximum given these observa-
tions, 𝑝 (𝑓 ∗ | D). The global optimum was discovered after 23 iterations,
somewhat slower than the alternatives described above. The value of
the optimum is known with almost complete confidence at termination.

7.7 multi-armed bandits and optimization

Several Bayesian optimization algorithms have been inspired by policies
for multi-armed bandits, a model system for sequential decision making multi-armed bandits

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 141

 common bayesian optimization policies

𝑝 (𝑓 ∗ | D)

Figure 7.16: The posterior after 10 (top) and 25 (bottom) steps of maximizing the mutual information between the observed
value 𝑦 and the location of the global maximum 𝑓 ∗ (7.17) on our running example. The tick marks show the
points chosen by the policy, progressing from top to bottom. Observations within 0.2 length scales of the
optimum are marked in red; the optimum was located on iteration 23.

under uncertainty. A multi-armed bandit problem can be interpreted as

a particular finite-dimensional analog of sequential optimization, and
effective algorithm design in both settings requires addressing many
shared concerns.
The classical multi-armed bandit problem considers a finite set of
set of arms, X “arms” X and an agent who must select a sequence of items from this
set. Selecting an arm 𝑥 results in a stochastic reward 𝑦 drawn from an
unknown distribution 𝑝 (𝑦 | 𝑥) associated with that arm; these rewards
are assumed to be independent of time and conditionally independent
given the chosen arm. The goal of the agent is to select a sequence of
Í
cumulative reward: § 6.2, p.114 arms {𝑥𝑖 } to maximize the cumulative reward (6.2) received, 𝑦𝑖 .
Multi-armed bandits have been studied as a model of many sequen-
tial decision processes arising in practice. For example, consider a doctor
caring for a series of patients with the same condition, who must deter-
mine which of two possible treatments is the best course of action. We
could model the sequence of treatment decisions as a two-armed bandit,
with patient outcomes determining the rewards. The Hippocratic oath
compels the doctor to discover the optimal arm (the best treatment) as
26 w. r. thompson (1933). On the Likelihood efficiently and confidently as possible to minimize patient harm, creating
that One Unknown Probability Exceeds An- a dilemma of effective assignment. In fact, this scenario of sequential clin-
other in View of the Evidence of Two Samples.
Biometrika 25(3–4):285–294.
ical trials was precisely the original motivation for studying multi-armed
bandits.26

142
 7.7. multi-armed bandits and optimization

To facilitate the following discussion, for each arm 𝑥 ∈ X, we define optimal policy with known rewards
𝜙 = 𝔼[𝑦 | 𝑥] to be its expected reward and will aggregate these into a
vector f when convenient. We also define

𝑥 ∗ ∈ arg max 𝔼[𝑦 | 𝑥] = arg max f; 𝑓 ∗ = max 𝔼[𝑦 | 𝑥] = max f

𝑥 ∈X 𝑥 ∈X

to be the index of an arm with maximal expected reward and the value of 27 All of the notation throughout this section is
that optimal reward, respectively.27 If the reward distributions associated chosen to align with that for optimization. In
a multi-armed bandit, an arm 𝑥 is associated
with each arm were known a priori, the optimal policy would be trivial: with expected reward 𝜙 = 𝔼[𝑦 | 𝑥 ]. In op-
we would always select the arm with the highest expected reward. This timization with zero-mean additive noise, a
policy generates expected reward 𝑓 ∗ in each iteration, and it is clear point 𝑥 is associated with expected observed
from linearity of expectation that this is optimal. Unfortunately, the value 𝜙 = 𝑓 (𝑥) = 𝔼[𝑦 | 𝑥 ].
reward distributions are unknown to the agent and must be learned from
observations instead. This complicates policy design considerably.
The only way we can learn about the reward distributions is to challenges in policy design
allocate resources to each arm and observe the outcomes. If the reward
distributions have considerable spread and/or overlap with each other,
a large number of observations may be necessary before the agent can
confidently conclude which arm is optimal. The agent thus faces an
exploration–exploitation dilemma, constantly forced to decide whether
to select an arm believed to have high expected reward (exploitation) 28 We will explore these connections in our dis-
or whether to sample an uncertain arm to better understand its reward cussion on theoretical convergence results for
Bayesian optimization algorithms in chapter
distribution (exploration). Ideally, the agent would have a policy that
10, p. 213.
efficiently explores the arms, so that in the limit of many decisions the
29 d. a. berry and b. fristedt (1985). Bandit
agent would eventually allocate an overwhelming majority of resources Problems: Sequential Allocation of Experiments.
to the best possible alternative. Chapman & Hall.
Dozens of policies for the multi-armed bandit problem have been pro- 30 s. bubeck and n. cesa-bianchi (2012). Re-
posed and studied from both the Bayesian and frequentist perspectives, gret Analysis of Stochastic and Nonstochas-
and many strong convergence results are known.28 Numerous variations tic Multi-armed Bandit Problems. Foundations
on the basic formulation outlined above have also received consideration and Trends in Machine Learning 5(1):1–122.
in the literature, and the interested reader may refer to one of several 31 t. lattimore and c. szepesvári (2020). Bandit
available exhaustive surveys for more information.29, 30, 31 Algorithms. Cambridge University Press.

The Bayesian optimal policy

A multi-armed bandit is fundamentally a sequential decision problem
under uncertainty, and we may derive an optimal expected-case policy
following our discussion in chapter 5. The selection of each arm is a
decision with action space X, and we must act under uncertainty about
the expected reward vector f. Over the course of 𝜏 decisions, we will
gather a dataset D𝜏 = (x𝜏 , y𝜏 ), seeking to maximize the cumulative
Í
reward (6.2): 𝑢 (D𝜏 ) = 𝑦𝑖 .
The key to the Bayesian approach is maintaining a belief about belief about expected rewards
the expected reward of each arm. We begin by choosing a prior over
the expected rewards, 𝑝 (f), and an observation model for the observed 32 This model is often conditionally indepen-
rewards given the index of an arm and its expected reward, 𝑝 (𝑦 | 𝑥, 𝜙).32 dent of the arm given the expected reward,
allowing the definition of a single observation
Now given an arbitrary set of previous observations D, we may derive a model 𝑝 (𝑦 | 𝜙).
posterior belief about the expected rewards, 𝑝 (f | D).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 143

 common bayesian optimization policies

optimal policy: § 5.2, p. 91 The optimal policy may now be derived following our previous
analysis. We make each decision by maximizing the expected reward by
termination, recursively assuming optimal future behavior (5.15–5.17).
Notably, the optimal decision for the last round is the arm maximizing
33 See § 7.10 for an analogous result in optimiza- the posterior mean reward, reflecting pure exploitation:33
tion.
𝑥𝜏 ∈ arg max 𝔼[f | D𝜏−1 ].
More exploratory behavior begins with the penultimate decision and
increases with the decision horizon.
running time of optimal policy: § 5.3, p. 99 Unfortunately, the cost of computing the optimal policy increases ex-
ponentially with the horizon. We must therefore find some mechanism to
design computationally efficient but empirically effective policies for use
in practice. This is precisely the same situation we face in optimization!

Optimization as an infinite-armed bandit

We may model continuous optimization as an infinite-armed bandit
34 r. agrawal (1995). The Continuum-Armed problem,34 and this analogy has proven fruitful. Suppose we seek to
Bandit Problem. siam Journal on Control and optimize an objective function 𝑓 : X → ℝ, where the domain X is now
Optimization 33(6):1926–1951.
infinite. We assume as usual that we can observe this function at any
point 𝑥 of our choosing, revealing an observation 𝑦 with distribution
𝑝 (𝑦 | 𝑥, 𝜙); 𝜙 = 𝑓 (𝑥). For the bandit analogy to be maximally appropriate,
we will further assume that the expected value of this observation is 𝜙:
𝔼[𝑦 | 𝜙] = 𝜙; this not unduly restrictive and is satisfied for example by
zero-mean additive noise.
Assuming an evaluation budget of 𝜏 observations, we may formulate
this scenario to a multi-armed bandit. We interpret each point 𝑥 ∈ X
as one of infinitely many arms, with each arm returning an expected
reward determined by the underlying objective function value 𝜙. With
some care, we may now adapt a multi-armed bandit policy to this setting.
correlation of rewards In the traditional multi-armed bandit problem, we assume that re-
wards are conditionally independent given the chosen arm index. As a
consequence, the reward from any selected arm provides no information
about the rewards of other arms. However, this independence would
render an infinite-armed bandit hopeless, as we would never be able to
determine the best arm with a finite budget. Instead, we must assume
that the rewards are correlated over the domain, so that each observation
35 In the multi-armed bandit literature, bandits can potentially inform us about the rewards of every other arm.35
with correlated rewards are known as restless This assumption is natural in optimization; the objective function
bandits, as our belief about arms may change
even when they are not selected (left alone):
must reflect some nontrivial structure, or optimization would also be
hopeless. In the Bayesian framework, we formalize our assumptions
p. whittle (1988). Restless Bandits: Activity regarding the structure of correlations between function values by choos-
Allocation in a Changing World. Journal of
Applied Probability 25(a):287–298 ing an appropriate prior distribution, which we may condition on avail-
able data to form the posterior belief, 𝑝 (𝑓 | D). In our bandit analogy,
this distribution encapsulates beliefs about the expected rewards of each
arm that can be used to derive effective policies.
cumulative vs. simple reward Why should we reduce optimization to the multi-armed bandit at
all? Notably, in optimization we are usually concerned with identifying

144
 7.8. maximizing a statistical upper bound

a single point in the domain maximizing the function, and variations

on the simple reward directly measure progress toward this end. It may
seem odd to focus on maximizing the cumulative reward, which judges a
dataset based on the average value observed rather than the maximum.
These aims are not necessarily incompatible. In the limit of many cumulative regret and convergence: § 10.1,
observations, we hope to guarantee the best arm will be eventually p. 214
identified so that we can guarantee convergence to optimal behavior.
Bandit algorithms are typically analyzed in terms of their cumulative 36 In the infinite-armed case, establishing the no-
regret, the difference between the cumulative reward received and that regret property (§ 10.1, p. 214) is not sufficient
to guarantee the global optimum will ever be
expected from the optimal policy. If this quantity decreases sufficiently evaluated exactly; however, we can conclude
quickly, we may conclude that the optimal arm is eventually identified that we evaluate points achieving objective
and selected. In our infinite-armed case, this implies the global optimum values within any desired tolerance of the max-
of the objective will eventually be located,36 suggesting the multi-armed imum.

bandit reduction is indeed reasonable.

7.8 maximizing a statistical upper bound

Effective strategies for both bandits and optimization require careful
consideration of the exploration–exploitation dilemma for success. Nu-
merous bandit algorithms have been built on the unifying principle of
optimism in the face of uncertainty,37 which has proven to be an effec- optimism in the face of uncertainty
tive heuristic for balancing these concerns. The key idea is to use any
available data to both estimate the expected reward of each arm and to 37 a. w. moore and c. g. atkeson (1993).
quantify the uncertainty in these estimates. When faced with a decision, Memory-based Reinforcement Learning:
Efficient Computation with Prioritized
we then always select the arm that would be optimal when allowing “the
Sweeping. neurips 1992.
benefit of the doubt”: the arm with the highest plausible expected reward
given the currently available information. Arms with highly uncertain
reward will have a correspondingly wide range of plausible values, and
this mechanism thus provides underexplored arms a so-called exploration
bonus38 commensurate with their uncertainty, encouraging exploration 38 r. s. sutton (1990). Integrated Architectures
of plausible optimal locations. for Learning, Planning, and Reacting Based on
Approximating Dynamic Programming. icml
To be more precise, assume we have gathered an arbitrary dataset 1990.
D, and consider an arbitrary point 𝑥. Consider the quantile function
associated with the predictive distribution 𝑝 (𝜙 | 𝑥, D):39 39 The quantile function satisfies the relation that
 𝜙 ≤ 𝑞 (𝜋 ; 𝑥, D) with probability 𝜋 .
𝑞(𝜋; 𝑥, D) = inf 𝜙 0 | Pr(𝜙 ≤ 𝜙 0 | 𝑥, D) ≥ 𝜋 .

We can interpret this function as a statistical upper confidence bound on upper confidence bound
𝜙: the value will exceed the bound only with tunable probability 1 − 𝜋.
As a function of 𝑥, we can interpret 𝑞(𝜋; 𝑥, D) as an optimistic esti-
mate of the entire objective function. The principle of optimism in the
face of uncertainty then suggests observing where this upper confidence
bound is maximized, yielding the acquisition function

𝛼 ucb (𝑥; D, 𝜋) = 𝑞(𝜋; 𝑥, D). (7.18)

Figure 7.17 shows upper confidence bounds for our example scenario example and discussion
corresponding to three values of the confidence parameter 𝜋. Unlike the
acquisition functions considered previously in this chapter, an upper

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 145

 common bayesian optimization policies

upper confidence bound, 𝛼 ucb (𝜋 = 0.8) next observation location

𝛼 ucb (𝜋 = 0.95)

𝛼 ucb (𝜋 = 0.999)

Figure 7.17: The upper confidence bound acquisition function (7.18) corresponding our running example for different
values of the confidence parameter 𝜋. The vertical axis for each acquisition function is shifted to the largest
observed function value. Increasing the confidence parameter leads to increasingly exploratory behavior.

confidence bound need not be nonnegative, so we shift the acquisition

function in these plots so that the best-seen function value intersects the
horizontal axis. We can see that relatively low confidence values (𝜋 = 0.8)
give little credit to locations with high uncertainty, and exploitation is
correspondence between certainty parameter heavily favored. By increasing this parameter, our actions reflect more
and exploration proclivity exploratory behavior.
In figure 7.18, we sequentially maximize the upper confidence bound
on our example function to select 20 observation locations using con-
fidence parameter 𝜋 = 0.999, corresponding to the bottom and most
exploratory example in figure 7.17. The global maximum was located
efficiently. Notably, the observation locations chosen in the early stages
of the search reflect more exploration than all the other methods we
have discussed thus far.
adjusting the exploration parameter Using an upper confidence bound policy in practice requires specify-
ing the exploration parameter 𝜋, and it may not always be clear how best
to do so. We face a similar challenge when choosing the improvement
target parameter for probability of improvement, and in fact this analogy
is sometimes remarkably intimate. For some models of the objective
40 d. r. jones (2001). A Taxonomy of Global Op- function, including Gaussian processes, a point maximizing the proba-
timization Methods Based on Response Sur- bility of improvement over a given threshold 𝜏 also maximizes an upper
faces. Journal of Global Optimization 21(4):345–
383
confidence bound for some confidence parameter 𝜋, and vice versa.40
Therefore the sets of points obtained by maximizing these acquisition

146
 7.8. maximizing a statistical upper bound

Figure 7.18: The posterior after 10 (top) and 20 (bottom) steps of the optimization policy induced by the upper confidence
bound acquisition function (7.2) on our running example. The confidence parameter was set to 𝜋 = 0.999. The
tick marks show the points chosen by the policy, progressing from top to bottom, during iterations 1–10 (top)
and 11–20 (bottom). Observations within 0.2 length scales of the optimum are marked in red; the optimum
was located on iteration 15.

functions over the range of their respective parameters are identical. We

will establish this relationship for Gaussian process models in the next equivalence of 𝛼 pi and 𝛼 ucb for gps: § 8.4,
chapter. p. 170
Little concrete advice is available for selecting the confidence param-
eter, as concerns such as model selection and calibration may have effects
on the upper confidence bound that are hard to foresee and account for.
Most authors select relatively large values in the approximate range
𝜋 ∈ (0.98, 1), with values of 𝜋 ≈ 0.999 being perhaps the most common.
In line with his advice regarding probability of improvement parameter
selection, jones suggests considering a wide range of confidence val- selecting improvement threshold: § 7.5, p. 134
ues,41 reflecting different exploration–exploitation tradeoffs. Figure 7.12
illustrates this concept by maximizing the probability of improvement 41 d. r. jones (2001). A Taxonomy of Global Op-
for a range of improvement thresholds; by the correspondence between timization Methods Based on Response Sur-
faces. Journal of Global Optimization 21(4):345–
these policies for Gaussian process models, this is also illustrative for 383.
maximizing upper confidence bounds.
The policy realized by sequentially maximizing upper confidence theoretical analysis: chapter 10, p. 213
bounds enjoys strong theoretical guarantees. For Gaussian process mod-
els, srinivas et al. proved this policy is guaranteed to effectively maxi- 42 n. srinivas et al. (2010). Gaussian Process Op-
mize the objective at a nontrivial rate under reasonable assumptions.42 timization in the Bandit Setting: No Regret
and Experimental Design. icml 2010.
One of these assumptions is that the confidence parameter must increase
asymptotically to 1 at a particular rate. Intuitively, the reason for this These results will be discussed at length in
growth is that our uncertainty in the objective function will typically chapter 10, p. 213.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 147

 common bayesian optimization policies

decrease as we continue to gather more data. As a result, we must simul-

taneously increase the confidence parameter to maintain a sufficient rate
of exploration. This idea of slowly increasing the confidence parameter
throughout optimization may be useful as a practical heuristic as well as
a theoretical device.

7.9 thompson sampling

In the early 20th century, thompson proposed a simple and effective
stochastic policy for the multi-armed bandit problem that has come to
43 w. r. thompson (1933). On the Likelihood be known as Thompson sampling.43 Faced with a set of alternatives, the
that One Unknown Probability Exceeds An- basic idea is to maintain a belief about which of these options is optimal
other in View of the Evidence of Two Samples.
Biometrika 25(3–4):285–294.
in light of available information. We then design each evaluation by sam-
pling from this distribution, yielding an adaptive stochastic policy. This
procedure elegantly addresses the exploration–exploitation dilemma:
sampling observations proportional to their probability of optimality
automatically encourages exploitation, while the inherent randomness
of the policy guarantees constant exploration. Thompson sampling can
be adopted from finite-armed bandits to continuous optimization, and
has enjoyed some interest in the Bayesian optimization literature.
definition Suppose we are at an arbitrary stage of optimization with data D.
The key object in Thompson sampling is the posterior distribution of
mutual information with 𝑥 ∗ : § 7.6, p. 139 the location of the global maximum 𝑥 ,∗ 𝑝 (𝑥 ∗ | D), a distribution we have
already encountered in our discussion on mutual information. Whereas
maximizing mutual information carefully maximizes the information we
expect to receive, Thompson sampling employs a considerably simpler
mechanism. We choose the next observation location by sampling from
this belief, yielding a nondeterministic optimization policy:
𝑥 ∼ 𝑝 (𝑥 ∗ | D). (7.19)
alternative interpretation: optimizing a At first glance, Thompson sampling appears fundamentally different
random acquisition function from the previous policies we have discussed, which were all defined in
terms of maximizing an acquisition function. However, we may resolve
this discrepancy while gaining some insight: Thompson sampling in fact
designs each observation by maximizing a random acquisition function.
The location of the global maximum 𝑥 ∗ is completely determined by
the objective function 𝑓, and we may exploit this relationship to yield
a simple two-stage implementation of Thompson sampling. We first
44 We give the sample a suggestive name! sample a random realization of the objective function from its posterior:44
𝛼 ts (𝑥; D) ∼ 𝑝 (𝑓 | D). (7.20)
We then optimize this function to yield the desired sample from 𝑝 (𝑥 ∗ | D),
our next observation location:
𝑥 ∈ arg max 𝛼 ts (𝑥 0; D).
𝑥 0 ∈X

From this point of view, we can interpret the sampled objective function
𝛼 ts as an ordinary acquisition function that is maximized as usual.

148
 7.9. thompson sampling

samples from 𝑝 (𝑓 | D)

sample from 𝑝 (𝑓 | D), 𝛼 ts next observation location

Figure 7.19: An illustration of Thompson sampling for our example optimization scenario. At the top, we show the
objective function posterior 𝑝 (𝑓 | D) and three samples from this belief. Thompson sampling selects the
next observation location by maximizing one of these samples. In the bottom three panels we show three
possible outcomes of this process, corresponding to each of the sampled objective functions illustrated in the
top panel.

Rather than representing an expected utility or a statistical upper optimal decision for random sample
bound, the acquisition function used in each round of Thompson sam-
pling is a hallucinated objective function that is plausible under our
belief. Whereas a Bayesian decision theoretic policy chooses the opti-
mal action in expectation while averaging over the uncertain objective
function, Thompson sampling chooses the optimal action for a randomly
sampled objective function. This interpretation of Thompson sampling is
illustrated for our example scenario in figure 7.19, showing three possible
outcomes for Thompson sampling. In this case, two samples would ex-
ploit the region surrounding the best-seen point, and one would explore
the region around the left-most observation.
Figure 7.20 shows the posterior belief of the objective after 15 rounds example and discussion
of Thompson sampling for our example scenario. The global maximum
was located remarkably quickly. Of course, as a stochastic policy, it is
not guaranteed that the behavior of Thompson sampling will resemble
this outcome. In fact, this was a remarkably lucky run! The most likely
locations of the optimum were ignored in the first rounds, quickly leading
to the discovery of the optimum on the left. Figure 7.21 shows another example of slow convergence

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 149

common bayesian optimization policies

Figure 7.20: The posterior after 15 steps of Thompson sampling (7.19) on our running example. The tick marks show the
points chosen by the policy, progressing from top to bottom. Observations within 0.2 length scales of the
optimum are marked in red; the optimum was located on iteration 8.

run of Thompson sampling on the same scenario. The global optimum

was not found nearly as quickly; however, it was eventually located after
approximately 80 iterations. In 100 repetitions of the policy varying the
random seed, the median iteration for discovering the optimum on this
example was 38, with only 12 seeds resulting in discovery in the first 20
iterations. Despite this sometimes slow convergence, the distribution
of the chosen evaluation locations nonetheless demonstrates continual
management of the exploration–exploitation tradeoff.

7.10 other ideas in policy construction

We have now discussed the two most pervasive approaches to policy
construction in Bayesian optimization: one-step lookahead and adapting
policies for multi-armed bandits. We have also introduced the most
popular Bayesian optimization policies encountered in the literature, all
of which stem from one of these two methods. However, there are some
additional ideas worthy of discussion.

Approximate dynamic programming beyond one-step lookahead

One-step lookahead offers tremendous computational benefits, but the
cost of these savings is extreme myopia in decision making. As we are
oblivious to anything that might happen beyond the present observation,
one-step lookahead can focus too much on exploitation. However, some
approximate dynamic programming: § 5.3, less myopic alternatives have been proposed based on more complex
p. 101 (and more costly!) approximations to the optimal policy.
The simplest idea in this direction is to extend the lookahead horizon,
two-step lookahead and the most tractable option is of course two-step lookahead. Given an
arbitrary one-step lookahead acquisition function 𝛼 (𝑥; D), the two-step
analog is (5.12):
h i
𝛼 2 (𝑥; D) = 𝛼 (𝑥; D) + 𝔼 max
0
𝛼 (𝑥 0
; D 0
) | 𝑥, D .
𝑥 ∈X

150
 7.10. other ideas in policy construction

Figure 7.21: The posterior after 80 steps of Thompson sampling (7.19) on our running example, using a random seed
different from figure 7.20. The global optimum was located on iteration 78.

Although this door is open for any of the decision-theoretic policies con-
sidered in this chapter, two-step expected improvement has received the
most attention. osborne et al. derived two-step expected improvement
and demonstrated good empirical performance on some test functions
compared with the one-step alternative.45 However, it is telling that they 45 m. a. osborne et al. (2009). Gaussian Processes
restricted their investigation to a limited number of functions due to the for Global Optimization. lion 3.
inherent computational expense. ginsbourger and le riche completed 46 d. ginsbourger and r. le riche (2010). To-
a contemporaneous exploration of two-step expected improvement and wards Gaussian Process-based Optimization
with Finite Time Horizon. moda 9.
provided an explicit example showing superior behavior from the less
myopic policy.46 Recently, several authors have revisited (2+)-step looka- 47 j. wu and p. i. frazier (2019). Practical
Two-Step Look-Ahead Bayesian Optimization.
head and developed sophisticated implementation schemes rendering neurips 2019
longer horizons more feasible.47, 48
48 s. jiang et al. (2020b). Efficient Nonmyopic
We provided an in-depth illustration and deconstruction of two-step Bayesian Optimization via One-Shot Multi-
expected improvement for our example scenario in figures 5.2–5.3. Note Step Trees. neurips 2020
that the two-step expected improvement is appreciable even for the
(useless!) options of evaluating at the previously observed locations, as
we can still make conscientious use of the following observation.
Figure 7.22 illustrates the progress of 20 evaluations designed by
maximizing two-step expected improvement for our example scenario.
Comparing with the one-step alternative in figure 7.4, the less myopic
policy exhibits somewhat more exploratory behavior and discovered the
optimum more efficiently – after 15 rather than 19 evaluations.
Rollout has also been considered as an approach to building nonmy- rollout: § 5.3, p. 102
opic optimization policies. Again the focus of these investigations has
been on expected improvement (or the related knowledge gradient), but
the underlying principles could be extended to other policies.
lam et al. combined expected improvement with several steps of 49 r. r. lam et al. (2016). Bayesian Optimization
rollout, again maximizing expected improvement as the base policy.49 The with a Finite Budget: An Approximate Dy-
namic Programming Approach. neurips 2016.
authors also proposed optionally adjusting the utility function through

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 151

 common bayesian optimization policies

Figure 7.22: The posterior after 10 (top) and 20 (bottom) steps of the optimization policy induced by maximizing two-step
expected improvement on our running example. The tick marks show the points chosen by the policy,
progressing from top to bottom, during iterations 1–10 (top) and 11–20 (bottom). Observations within 0.2
length scales of the optimum are marked in red; the optimum was located on iteration 15.

multiplicative discounting to encourage earlier rather than later progress

50 Such a discount factor is common in infinite- during the rollout steps.50 For some combinations of the rollout horizon
horizon decision problems: and the discount factor, the resulting policies outperformed common
d. p. bertsekas (2017). Dynamic Programming one-step lookahead policies on a suite of synthetic test functions. yue
and Optimal Control. Vol. 1. Athena Scientific. and al kontar described a mechanism for dynamically choosing the
51 x. yue and r. al kontar (2020). Why Non- rollout horizon based on the potential impact of a misspecified model,51
myopic Bayesian Optimization is Promising exactly the issue that gave kushner pause.52
and How Far Should We Look-ahead? A Study The additional computational burden of rollout limited lam et al. to
via Rollout. aistats 2020.
a relatively short rollout horizon on the order of 4–5. Although consid-
52 See p. 125.
erably less myopic than one-step lookahead, the true decision horizon
can be much greater, especially during the early stages of optimization.
batch rollout: § 5.3, p. 103 gonzález et al. proposed an alternative approach based on batch rollout
that can effectively look farther ahead at the expense of ignoring de-
53 j. gonzález et al. (2016b). glasses: Relieving pendence among future decisions.53 The algorithm – dubbed glasses by
The Myopia Of Bayesian Optimization. ais- the authors as it counteracts myopia – augments expected improvement
tats 2016.
with a single rollout step that designs a batch of additional observa-
54 The policy used in glasses is described in tion locations of size equal to the remaining evaluation budget.54 These
j. gonzález et al. (2016a). Batch Bayesian Opti- points serve as a rough simulation of the decisions that might follow
mization via Local Penalization. aistats 2016, after making a proposed observation. A potential observation location is
then evaluated by the expected improvement gained by simultaneously
as well as in § 11.3, p. 257; however, any desired
alternative could also be used. evaluating at that point as well as the constructed batch, realizing an
efficient and budget-aware nonmyopic policy. glasses outperformed
myopic and nonmyopic baselines on synthetic test functions, and the

152
 7.10. other ideas in policy construction

expected improvement, 𝛼 ei next observation location

modified expected improvement, 𝛼 ei

0 (𝜀 = 1)

0 (𝜀 = 10)
𝛼 ei

Figure 7.23: The modified expected improvement acquisition function (7.21) for our running example for different values
of the target improvement 𝜀. The target is expressed as a fraction of the range of the posterior mean over the
space. Increasing the target improvement leads to increasingly exploratory behavior.

authors also demonstrated that dynamically setting the batch size to the
remaining budget outperformed an arbitrary fixed size, suggesting that
budget adaptation was important for success.
jiang et al. continued this thread with an even more dramatic ap-
proximation dubbed binoculars, potentially initiating an arms race
toward increasingly nonmyopic acronyms.55 The idea is to construct a 55 s. jiang et al. (2020a). binoculars for Ef-
single batch observation in each iteration, then select a point from this ficient, Nonmyopic Sequential Experimental
Design. icml 2020.
batch for evaluation. This represents an extreme computational savings
over glasses, which must construct a batch anew for every proposed
observation location. However, the method retains the same fundamental
motivation: well-designed batch policies automatically induce diversity batch observations: § 11.3, p. 252
among batch members, encouraging exploration in the resulting sequen-
tial policy (see figure 11.3). The strong connection between the optimal
batch and sequential policies (11.9–11.10) provides further motivation for
this approach. jiang et al. also conducted a study of optimization per-
formance versus the cost of computing the policy: one-step lookahead,
binoculars, glasses, and rollout comprised the Pareto frontier, with Pareto frontier: § 11.7, p. 269
each method increasing computational effort by an order of magnitude.

Artificially encouraging exploration

Another more indirect approach to nonmyopic policy design is to modify
a one-step lookahead acquisition function to artificially encourage more

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 153

 common bayesian optimization policies

Figure 7.24: The posterior after 15 steps of the streltsov and vakili optimization policy on our running example. The
tick marks show the points chosen by the policy, progressing from top to bottom. Observations within 0.2
length scales of the optimum are marked in red; the optimum was located on iteration 14.

exploratory behavior when the remaining budget is significant. This can

be motivated by the nature of the optimal policy, which maximizes a
combination of immediate reward (exploitation) with expected future
reward (exploration) (5.18). As the decision horizon increases, the explo-
ration term may become increasingly dominant in this score, suggesting
that optimal behavior entails early exploration of the domain followed
by later exploitation and refinement.
Both the probability of improvement (7.5) and upper confidence
bound acquisition functions already feature a parameter controlling the
exploration–exploitation tradeoff, which can be dynamically maintained
throughout optimization. This idea is quite old, reaching back to the
earliest papers on Bayesian optimization. kushner for example provided
detailed advice on adjusting the threshold in probability of improvement
to transition from early exploration to increasing levels of exploitation
56 h. j. kushner (1964). A New Method of Locat- when appropriate.56
ing the Maximum Point of an Arbitrary Multi- In the case of exact observation, it is also possible to modify ex-
peak Curve in the Presence of Noise. Journal
of Basic Engineering 86(1):97–106.
pected improvement (7.3) to incorporate a similar parameter. Rather
than measuring improvement with respect to the utility of the current
data 𝑢 (D) = 𝜙 ,∗ we measure with respect to an inflated value 𝜙 ∗ + 𝜀, with
no credit given for improvements less than this amount. The result is a
modified expected improvement acquisition function:
 
0
𝛼 ei (𝑥; D, 𝜀) = 𝔼 max{𝜙 − [𝜙 ∗ + 𝜀], 0} | 𝑥, D . (7.21)

As with probability of improvement, larger improvement thresholds en-

courage increasing exploration, as illustrated in figure 7.23. For extreme
values, the modified expected improvement drops to effectively zero
except in regions with significant uncertainty. It is not obvious how this
idea can be extended handle noisy observations, as the simple reward of
the updated dataset may in fact decrease, raising the question of how to
correctly define “sufficient improvement.”
mockus proposed a scheme to set the threshold for Gaussian pro-
cess models with constant mean and stationary covariance where the

154
 7.10. other ideas in policy construction

threshold was set dynamically based on the remaining budget, using the
asymptotic behavior of the maximum of iid Gaussian random variables.57 57 j. mockus (1989). Bayesian Approach to Global
This can be interpreted as an approximate batch rollout policy where re- Optimization: Theory and Applications. Kluwer
Academic Publishers. [§ 2.5]
maining decisions are simulated by fictitious uncorrelated observations;
for some models this serves as an efficient simulation of random rollout.
The modified expected improvement (7.21) was also the basis for an
unusual policy proposed by streltsov and vakili.58 Let 𝑐 : X → ℝ >0 58 s. streltsov and p. vakili (1999). A Non-
quantify the cost of making an observation at any proposed location; in myopic Utility Function for Statistical Global
Optimization Algorithms. Journal of Global
the simplest case we could take the cost to be constant. To evaluate the Optimization 14(3):283–298.
promise of making an observation at 𝑥, we solve the equation
0
𝛼 ei (𝑥; D, 𝛼 sv ) = 𝑐 (𝑥)
for 𝛼 sv , which will serve as the acquisition function value at 𝑥.59 That is, 59 As 𝛼 ei
0 is monotonically decreasing with re-

we solve for the improvement threshold that would render an observa- spect to 𝛼 sv and approaches zero as 𝛼 sv → ∞,
the unique solution can be found efficiently
tion at 𝑥 cost-prohibitive in expectation, and design each observation via bisection.
to coincide with the last point to be ruled out when considering in-
creasingly demanding thresholds. The resulting policy shows interesting
behavior, at least on our running example; see figure 7.24. After effec-
tive initial exploration, the global optimum was located on iteration
14. The behavior is similar to the upper confidence bound approach in
figure 7.18, and indeed streltsov and vakili showed that the proposed
method can be understood as a variation on this method with a location-
dependent upper confidence quantile depending on observation cost and
uncertainty.
An even simpler mechanism for injecting exploration into an ex-
isting policy is to occasionally make decisions randomly or via some
other policy encouraging pure exploration. de ath et al. for example
considered a family of 𝜀-greedy policies where a one-step lookahead
policy is interrupted, with probability 𝜀 in each iteration, by evaluating at 60 g. de ath et al. (2021). Greed is Good: Explo-
a location chosen uniformly at random from the domain.60 These policies ration and Exploitation Trade-offs in Bayesian
Optimisation. acm Transactions on Evolution-
delivered impressive empirical performance, even when the one-step ary Learning and Optimization 1(1):1–22.
lookahead policy was as simple as maximizing the posterior mean (see
below).

Lookahead for cumulative reward?

Notably missing from our discussion on one-step lookahead policies cumulative reward: § 6.2, p.114
was the cumulative reward utility function. Unfortunately, in this case
one-step lookahead does not produce a particularly useful optimization
policy. Suppose we adopt the cumulative reward utility function (6.7),
Í
𝑢 (D) = 𝑦𝑖 . Then marginal gain in utility from a measurement at 𝑥 is
simply the observed value 𝑦. Therefore the expected one-step marginal
gain is the posterior predictive mean:
𝛼 (𝑥; D) = 𝔼[𝑦 | 𝑥, D],
which for zero-mean additive noise reduces to the posterior mean of 𝑓 : posterior mean acquisition function

𝛼 (𝑥; D) = 𝜇D (𝑥).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 155

common bayesian optimization policies

Figure 7.25: The posterior after 15 steps of two-step lookahead for cumulative reward (7.22) on our running example. The
tick marks show the points chosen by the policy, progressing from top to bottom. The policy becomes stuck
on iteration 7.

From cursory inspection of our example scenario in figure 7.1, we can see
that maximizing this acquisition function can cause the policy to become
becoming “stuck” “stuck” with no chance of recovery. In our example, the posterior mean
is maximized at the previously best-seen point, so the policy will select
this point forevermore.
two-step lookahead for cumulative reward It is also interesting to consider two-step lookahead, where the ac-
quisition function becomes (5.12):
h i
𝛼 2 (𝑥; D) = 𝜇D (𝑥) + 𝔼 max
0
𝜇 D
0
0 (𝑥 ) | 𝑥, D .
𝑥 ∈X

After subtracting a constant (the global reward of the current data), we

have
𝛼 2 (𝑥; D) = 𝜇D (𝑥) + 𝛼 kg (𝑥; D), (7.22)
the sum of the posterior mean and the knowledge gradient (7.4), reflecting
both exploitation and exploration. Unfortunately even this less myopic
option can become stuck due to overexploitation, as illustrated in Figure
7.25.

7.11 summary of major ideas

Although we have covered a lot of ground in this chapter, there were
one-step lookahead: § 7.2, p. 124 really only two big ideas in policy construction: one-step lookahead and
multi-armed bandits: § 7.7, p. 141 adopting successful policies from multi-armed bandits. However, there
remains significant opportunity for novelty in this space.

156
 8
COMP U TING POLICIES WI TH GAUSSIAN PROCESSES

In the last chapter we introduced several notable Bayesian optimization

policies in a model-agnostic setting, concentrating on their motivation
and behavior while ignoring computational details. In this chapter we
will provide further information for effectively implementing these poli-
cies. We will focus on Gaussian process models of the objective function, Gaussian processes: chapter 2, p. 15
combined with either an exact or additive Gaussian noise observation observation models: § 1.1, p. 4
model; this family accounts for the vast majority of models encountered
in practice. However, we will discuss computation for alternative model alternative models: § 8.11, p. 196
classes such as Bayesian neural networks at the end of the chapter.
Implementing each policy in the previous chapter ultimately requires 1 Even the nondeterministic Thompson sam-
optimizing some acquisition function over the domain:1 pling policy, which may be realized by opti-
mizing a random acquisition function: § 7.9,
𝑥 ∈ arg max 𝛼 (𝑥 0; D). p. 148.
𝑥 0 ∈X

We will demonstrate how to compute (or approximate) each of these

acquisition functions with respect to Gaussian process models. Some
will admit exact analytical expressions; when this is not possible, we will
describe effective approximation schemes. In Euclidean domains, we will
also show how to compute the gradient of these acquisition functions
with respect to the proposed observation location, allowing efficient
optimization via gradient methods. These gradient computations will
sometimes be somewhat involved (but not difficult), and we will defer gradients of common acquisition functions:
some details to an appendix for the sake of brevity. § c.3, p. 308
The order of our presentation will differ from that in the previous
chapter. Here we will begin with the acquisition functions for which
exact computation is possible, then develop approximation techniques
for those that remain. First, we pause to establish notation for important
reoccurring quantities.

8.1 notation for objective function model

As usual, let us consider an objective function 𝑓 : X → ℝ, with a Gaus-
sian process belief conditioned on arbitrary observations D = (x, y): Gaussian process on 𝑓, GP (𝑓 ; 𝜇D , 𝐾D )

𝑝 (𝑓 | D) = GP (𝑓 ; 𝜇D , 𝐾D ). (8.1)

Our main task in this chapter will be to compute a given acquisition

function at an arbitrary location 𝑥 ∈ X with respect to this belief. Given a
proposed location 𝑥, we will write the predictive distribution for 𝜙 = 𝑓 (𝑥) predictive distribution for 𝜙, N (𝜙; 𝜇, 𝜎 2 )
as:
𝑝 (𝜙 | 𝑥, D) = N (𝜙; 𝜇, 𝜎 2 ), (8.2)
where the predictive mean and variance predictive mean and variance for 𝜙: 𝜇, 𝜎 2

𝜇 = 𝜇D (𝑥); 𝜎 2 = 𝐾D (𝑥, 𝑥) (8.3)

depend implicitly on 𝑥. We will always treat 𝑥 as given and fixed, so this

convention will not lead to ambiguity.

This material will be published by Cambridge University Press as Bayesian Optimization. 157
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
computing policies with gaussian processes

We will also require the predictive distribution for the observed value
𝑦 resulting from a measurement at 𝑥. In addition to the straightforward
case of exact measurements, where 𝑦 = 𝜙 and the predictive distribution
is given above (8.2), we will also consider corruption by independent,
zero-mean additive Gaussian noise:

𝑝 (𝑦 | 𝜙, 𝜎𝑛 ) = N (𝑦; 𝜙, 𝜎𝑛2 ). (8.4)

observation noise scale, 𝜎𝑛 Again we allow the noise scale 𝜎𝑛 to depend on 𝑥 if desired. We will
predictive distribution for 𝑦, N (𝑦; 𝜇, 𝑠 2 ) notate the resulting predictive distribution for 𝑦 with:

𝑝 (𝑦 | 𝑥, D, 𝜎𝑛 ) = N (𝑦; 𝜇, 𝜎 2 + 𝜎𝑛2 ) = N (𝑦; 𝜇, 𝑠 2 ), (8.5)

where 𝜇 and 𝜎 2 are the predictive moments of 𝜙 (8.3) and 𝑠 2 = 𝜎 2 + 𝜎𝑛2 is

predictive variance for 𝑦, 𝑠 2 the predictive variance for 𝑦, which again depends implicitly on 𝑥. When
no distinction between exact and noisy observations is necessary, we
will use the above general notation (8.5). With exact observations, the
observation noise scale is identically zero, and we have 𝑠 2 = 𝜎 2.
prime notation for post-observation We will retain our convention from last chapter of indicating quanti-
quantities ties available after acquiring a proposed observation with a prime symbol.
For example x0 = x ∪ {𝑥 } represents the updated set of observation loca-
tions after adding an observation at 𝑥, and D 0 = (x,0 y0) represents the
current data augmented with the observation (𝑥, 𝑦) – a random variable.
general form of gradient and dependence on The value of an acquisition function 𝛼 at a point 𝑥 will naturally
parameter gradients depend on the distribution of the corresponding observation 𝑦, and
its gradient must reflect this dependence. Applying the chain rule, the
general form of the gradient will be written in terms of the gradient of
the predictive parameters:

𝜕𝛼 𝜕𝛼 𝜕𝜇 𝜕𝛼 𝜕𝑠
= + . (8.6)
𝜕𝑥 𝜕𝜇 𝜕𝑥 𝜕𝑠 𝜕𝑥

The gradients of the predictive mean and standard deviation for a Gaus-
sian process are easily computable assuming the prior mean and covari-
gradients of gp predictive distribution: § c.2, ance functions and the observation noise scale are differentiable, and
p. 308 general expressions are provided in an appendix.

8.2 expected improvement

expected improvement: § 7.3, p. 127 We first consider the expected improvement acquisition function (7.2),
simple reward: § 6.1, p. 112 the expected marginal gain in simple reward (6.3):
 
𝛼 ei (𝑥; D) = 𝔼 max 𝜇D0 (x0) | 𝑥, D − max 𝜇D (x). (8.7)

Remarkably, this expectation can be computed analytically for Gaussian

processes with both exact and noisy observations. We will consider each
case separately, as the former is considerably simpler and the latter
involves minor controversy.

158
 8.2. expected improvement

𝑝 (𝜙 | 𝑥, D) Figure 8.1: The expectation required to compute noiseless ex-

max(𝜙 − 𝜙 ,∗ 0) pected improvement: linear improvement for val-
ues exceeding the incumbent 𝜙 ∗, integrated against
a Gaussian distribution (8.2).
𝜙
𝜙∗

Expected improvement without noise

Expected improvement assumes a convenient form when measurements
are exact (7.3):
∫
𝛼 ei (𝑥; D) = max(𝜙 − 𝜙 ,∗ 0) N (𝜙; 𝜇, 𝜎 2 ) d𝜙 . (8.8)

Here 𝜙 ∗ is the previously best seen, incumbent objective function value, incumbent function value, 𝜙 ∗
and max(𝜙 − 𝜙 ,∗ 0) measures the improvement offered by observing a
value of 𝜙. Figure 8.1 illustrates this integral.
To proceed, we resolve the max operator to yield two integrals:
∫∞ ∫∞
𝛼 ei (𝑥; D) = 𝜙 N (𝜙; 𝜇, 𝜎 2 ) d𝜙 − 𝜙 ∗ N (𝜙; 𝜇, 𝜎 2 ) d𝜙,
𝜙∗ 𝜙∗

both of which can be computed easily assuming 𝜎 > 0.2 The first term is 2 In the degenerate case 𝜎 = 0, we simply have
proportional to the expected value of a normal distribution truncated at 𝛼 ei (𝑥; D) = max(𝜇 − 𝜙 ,∗ 0).
𝜙 ∗, and the second term is the complementary normal cdf scaled by 𝜙 ∗.
The resulting acquisition function can be written conveniently in terms
of the standard normal pdf and cdf:
 𝜇 − 𝜙∗   𝜇 − 𝜙∗ 
𝛼 ei (𝑥; D) = (𝜇 − 𝜙 ∗ ) Φ + 𝜎𝜙 . (8.9)
𝜎 𝜎
Examining this expression, it is tempting to interpret its two terms exploitation and exploration
as respectively encouraging exploitation (favoring points with high
expected value 𝜇) and exploration (favoring points with high uncertainty partial derivatives with respect to predictive
𝜎). Indeed, taking partial derivatives with respect to 𝜇 and 𝜎, we have: distribution parameters

𝜕𝛼 ei  𝜇 − 𝜙∗  𝜕𝛼 ei  𝜇 − 𝜙∗ 
=Φ > 0; =𝜙 > 0.
𝜕𝜇 𝜎 𝜕𝜎 𝜎
Expected improvement is thus monotonically increasing in both 𝜇 and
𝜎. Increasing a point’s expected value naturally makes the point more
favorable for exploitation, and increasing its uncertainty makes it more
favorable for exploration. Either action would increase the expected
improvement. The tradeoff between these two concerns is considered
automatically and is reflected in the magnitude of the derivatives above.
Maximization of expected improvement in Euclidean domains may gradient of expected improvement without
be guided by its gradient with respect to the proposed evaluation location noise
𝑥. Using the results above and applying the chain rule, we have (8.6):
𝜕𝛼 ei  𝜇 − 𝜙 ∗  𝜕𝜇  𝜇 − 𝜙 ∗  𝜕𝜎
=Φ +𝜙 ,
𝜕𝑥 𝜎 𝜕𝑥 𝜎 𝜕𝑥

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 159

 computing policies with gaussian processes

which expresses the gradient in terms of the implicit exploration–exploitation

tradeoff parameter and the change in the predictive distribution.

Expected improvement with noise

Computing expected improvement with noisy observations is somewhat
more complicated than in the noiseless case. In fact, there is not even
universal agreement regarding what the definition of noisy expected
improvement should be! The situation is so nuanced that jones et al.
sidestepped the issue entirely in their landmark paper:3

3 d. r. jones et al. (1998). Efficient Global Opti- Unfortunately it is not immediately clear how to extend our opti-
mization of Expensive Black-Box Functions. mization algorithm to the case of noisy functions. With noisy data,
Journal of Global Optimization 13(4):455–492.
we really want to find the point where the signal is optimized. Sim-
ilarly, our expected improvement criterion should be defined in
terms of the signal component. [emphasis added by jones et al.]

This summarizes the main challenge to optimization with noisy observa-

tions: how can we determine whether a particularly high observed value
reflects a true underlying effect or is merely an artifact of noise? Several
alternative definitions and heuristics have been proposed to address this
question, which we will discuss further shortly.
We will first argue that seeking to maximize the simple reward utility
(6.3) is precisely aligned with the goal outlined by jones et al. With ap-
propriate modeling of observation noise, the objective function posterior
exactly represents our belief about the “signal component”: the latent
objective 𝑓. Simple reward evaluates progress directly with respect to
this belief, only ascribing merit to observations that improve the maxi-
mum of the posterior mean and thus the expected outcome of an optimal
terminal recommendation. Notably, an excessively noisy observation
has weak correlation with the underlying objective function value and
behavior of posterior moments: § 2.2, p. 21 thus yields little change in the posterior mean (2.12). Therefore even an
extremely high outcome would produce only a minor improvement to
the simple reward. As a result, the expected improvement of such a point
would be relatively small, exactly the desired behavior.
any point may maximize the posterior mean Unfortunately, observation noise renders expected improvement
somewhat more complicated than the exact case, due to the nature of the
updated simple reward. After an exact observation, the simple reward
can only be achieved at one of two locations: either the observed point
or the incumbent. However, with noisy observations, inherent uncer-
tainty in the objective function implies that the updated simple reward
could be achieved anywhere, including a point that previously appeared
suboptimal. We must account for this possibility in the computation,
increasing its complexity. Fortunately, exact computation is still possible
for Gaussian processes and additive Gaussian noise by adopting a proce-
4 p. frazier et al. (2009). The Knowledge- dure originally described by frazier et al. for computing the knowledge
Gradient Policy for Correlated Normal Beliefs. gradient on a discrete domain;4 we will return to this method in our
informs Journal on Computing 21(4):599–613.
discussion on knowledge gradient in the next section.

160
 8.2. expected improvement

If we define 𝜇 ∗ = max 𝜇 D (x) to represent the simple reward of the current simple reward, 𝜇 ∗
current data, then we must compute:
∫
 
𝛼 ei (𝑥; D) = max 𝜇D0 (x0) − 𝜇 ∗ N (𝑦; 𝜇, 𝑠 2 ) d𝑦.

We first reduce this computation to an expectation of the general form reduction to evaluation of 𝑔 (a, b)
∫
𝑔(a, b) = max(a + b𝑧) 𝜙 (𝑧) d𝑧, (8.10)

where a, b ∈ ℝ𝑛 are arbitrary vectors and 𝑧 is a standard normal random 5 The dimensions of a and b can be arbitrary as
variable.5 Note that given the observation 𝑦, the updated posterior mean long as they are equal.
at x0 is a vector we may compute in closed form (2.19):
𝐾D (x,0 𝑥) 𝑦 − 𝜇
𝜇D0 (x0) = 𝜇 D (x0) + .
𝑠 𝑠
This update is linear in 𝑦. Applying the transformation 𝑦 = 𝜇 + 𝑠𝑧 yields

𝜇D0 (x0) = a + b𝑧, (8.11)

where definition of a, b
𝐾D (x,0 𝑥)
a = 𝜇D (x0); b= , (8.12)
𝑠
and we may express expected improvement in the desired form:

𝛼 ei (𝑥; D) = 𝑔(a, b) − 𝜇 ∗. (8.13)

As a function of 𝑧, a + b𝑧 is a set of lines with intercepts and slopes geometric intuition of 𝑔

given by the entries of the a and b vectors, respectively. In the context
of expected improvement, these lines represent the updated posterior
mean values for each point of interest x0 as a function of the 𝑧-score
of the noisy observation 𝑦. See figure 8.2 for an illustration. Note that
the points with the highest correlation with the proposed point have
the greatest slope (8.12), as our belief at these locations will be strongly
affected by the outcome.
Now max(a + b𝑧) is the upper envelope of these lines, a convex
piecewise linear function shown in figure 8.3. The interpretation of this
envelope is the simple reward of the updated dataset given the 𝑧-score
of 𝑦, which will be achieved at some point in x.0 For this example, the up-
dated posterior mean could be maximized at one of four locations: either
at one of the points on the far right given a relatively high observation,
or at a backup point farther left given a relatively low observation. Note
that in the latter case the simple reward will decrease.6 6 However, the expected marginal gain is always
With this geometric intuition in mind, we can deduce that 𝑔 is invari- positive.
ant to transformations that do not alter the upper envelope. In particular,
𝑔 is invariant both to reordering the lines by applying an identical per- invariance to permutation
mutation to a and b and also to the deletion of lines that never dominate.
In the interest of notational simplicity, we will take advantage of these in- invariance to deletion of always dominated
variances and only consider evaluating 𝑔 when every given line achieves lines

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 161

 computing policies with gaussian processes

𝑝 (𝑧)
a + b𝑧

𝑧
𝑥

Figure 8.2: The geometric intuition of the 𝑔(a, b) function. Left: if we make a measurement at 𝑥, the 𝑧-score of the observed
value completely determines the updated posterior mean at that point and all previously observed points.
Right: as a function of 𝑧, the updated posterior mean at each of these points is linear; here the color of each
line corresponds to the matching point on the left. The slope and intercept of each line can be determined
from the posterior (8.12). Not all lines are visible.

maximal value on some interval and the lines appear in strictly increasing
order of slope:
𝑏 1 < 𝑏 2 < · · · < 𝑏𝑛 .
7 Briefly, we sort the lines in ascending order frazier et al. give a simple and efficient algorithm to process a set of
of slope, then add each line in turn to a set 𝑛 lines to eliminate any always-dominated lines, reorder the remainder in
of dominating lines, checking whether any
previously added lines need to be removed
increasing slope, and identify their intervals of dominance in O(𝑛 log 𝑛)
and updating the intervals of dominance. time.7 The output of this procedure is a permutation matrix P, possibly
with some rows deleted, such that

𝑔(a, b) = 𝑔(Pa, Pb) = 𝑔(𝜶 , 𝜷), (8.14)

and the new inputs (𝜶 , 𝜷) satisfy the desired properties. We will assume
below that the inputs have been preprocessed in such a manner.
Given a set of lines in the desired form, we may partition the real
line into a collection of 𝑛 intervals

(−∞ = 𝑐 1, 𝑐 2 ) ∪ (𝑐 2, 𝑐 3 ) ∪ · · · ∪ (𝑐𝑛 , 𝑐𝑛+1 = +∞), (8.15)

8 Reordering the lines in order of increasing such that the 𝑖th line 𝑎𝑖 + 𝑏𝑖 𝑧 dominates on the corresponding interval
slope guarantees this correspondence: the line (𝑐𝑖 , 𝑐𝑖+1 ).8 This allows us to decompose the desired expectation (8.10)
with minimal slope is always the “leftmost” in
the upper envelope, etc.
into a sum of contributions on each interval:
∑︁ ∫ 𝑐𝑖+1
𝑔(a, b) = (𝑎𝑖 + 𝑏𝑖 𝑧) 𝜙 (𝑧) d𝑧.
𝑖 𝑐𝑖

Finally, we may compute each integral in the sum in closed form:

∑︁    
𝑔(a, b) = 𝑎𝑖 Φ(𝑐𝑖+1 ) − Φ(𝑐𝑖 ) + 𝑏𝑖 𝜙 (𝑐𝑖 ) − 𝜙 (𝑐𝑖+1 ) , (8.16)
𝑖

162
 8.2. expected improvement

𝑝 (𝑧)
max(a + b𝑧)

𝑧
𝑥

Figure 8.3: After a measurement at 𝑥, the updated simple reward can be achieved at one of four points (left), whose
corresponding lines comprise the upper envelope max(a + b𝑧) (right). The colors of the line segments on the
right correspond to the possible updated maximum locations on the left. The lightest point on the left serves
as a “backup option” if the observed value is low.

allowing efficient and exact computation of expected improvement in the

noisy case. The main bottleneck is the modest O(𝑛 log 𝑛) preprocessing
step required.
This expression reverts to that for exact measurements (8.9) in the compatibility with noiseless case
absence of noise. In that case, we have b = [0, 𝑠] > , and the upper envelope
only contains lines corresponding to the incumbent and newly observed
point, which intersect at the incumbent value 𝑐 1 = 𝜙 .∗ Geometrically, the
situation collapses to that in figure 8.1, and it is easy to confirm (8.9) and
(8.16) coincide.
Although tedious, we may compute the gradient of 𝑔 and thus the gradient of noisy expected improvement:
gradient of expected improvement (8.13); the details are in an appendix. § c.3, p. 308

Alternative formulations of noisy expected improvement

Although thematically consistent and mathematically straightforward,
our approach of computing expected improvement as the expected
marginal gain in simple reward is not common and may be unfamil-
iar to some readers.
Over the years, numerous authors have grappled with the best defi-
nition of noisy expected improvement. A typical approach is to begin
with the convenient formula (8.9) in the noiseless regime, then work
“backwards” by identifying potential issues with its application to noisy
data and suggesting a heuristic correction. This strategy of “fixing” exact
expected improvement is in opposition to our ground-up approach of
first defining a well-grounded utility function and only then working
out the expected marginal gain. picheny et al. provided a survey of such 9 v. picheny et al. (2013b). A benchmark of
approximations – representing a total of eleven different acquisition kriging-based infill criteria for noisy optimiza-
tion. Structural and Multidisciplinary Opti-
strategies – accompanied by a thorough empirical investigation.9 No- mization 48(3):607–626.
tably, the authors were familiar with the exact computation we outlined
in the previous section and recommended it for use in practice.
One popular idea in this direction is to take the formula from the expected improvement with plug-in
exact case (8.9) and substitute a plug-in estimate for the now unknown estimator of 𝜙 ∗

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 163

 computing policies with gaussian processes

plug-in estimate, 𝜙 ∗ ≈ max y (8.17) next observation location

plug-in estimate, 𝜙 ∗ ≈ max 𝜇D (x) (8.18)

expected improvement (8.16)

Figure 8.4: Expected improvement using different plug-in estimators (8.17–8.18) compared with the noisy expected
improvement as the expected marginal gain in simple reward (8.7).

incumbent value 𝜙 ∗. Several possibilities for this estimate have been put
forward. One option is to plug in the maximum noisy observation:
𝜙 ∗ ≈ max y. (8.17)
maximum noisy value “utility” function: § 6.1, However, this may not always behave as expected for the same reason
p. 113 the maximum observed value does not serve as a sensible utility function
(6.6). With very noisy data, the maximum observed value is most likely
inflating expected improvement threshold: spurious rather than a meaningful goalpost. Further, as we are likely
§ 7.10, p. 154 to overestimate our progress due to bias in this estimate, the resulting
behavior may become excessively exploratory. The approximation will
eventually devolve to expected improvement against an inflated thresh-
old (7.21), which may overly encourage exploration; see figure 7.23. An
especially spurious observation can bias the estimate in (8.17) (and our
behavior) for a considerable time.
Opinions on the proposed approximation using this simple plug-in
10 v. picheny et al. (2013b). A benchmark of
kriging-based infill criteria for noisy optimiza- estimate (8.17) vary dramatically. picheny et al. discarded the idea out
tion. Structural and Multidisciplinary Opti- of hand as “naïve” and lacking robustness.10 The authors also found
mization 48(3):607–626. it empirically inferior in their investigation and described the same
11 v. nguyen et al. (2017). Regret for Expected over-exploratory effect and explanation given above. On the other hand,
Improvement over the Best-Observed Value nguyen et al. described the estimator as “standard” and concluded it was
and Stopping Condition. acml 2017.
empirically preferable!11

164
 8.2. expected improvement

An alternative estimator is the simple reward of the data (6.3) :12, 13 12 e. vazqez et al. (2008). Global optimization
based on noisy evaluations: an empirical study
𝜙 ∗ ≈ max 𝜇D (x), (8.18) of two statistical approaches. icipe 2008.
13 z. wang and n. de freitas (2014). Theoretical
which is less biased and may be preferable. A simple extension is to Analysis of Bayesian Optimization with Un-
maximize other predictive quantiles,10, 14 and huang et al. recommend known Gaussian Process Hyper-Parameters.
arXiv: 1406.7758 [stat.ML].
using a relatively low quantile, specifically Φ(−1) ≈ 0.16, in the interest
14 d. huang et al. (2006b). Global Optimization
of risk aversion. of Stochastic Black-Box Systems via Sequen-
In figure 8.4 we compare noisy expected improvement with two tial Kriging Meta-Models. Journal of Global
plug-in approximations. The plug-in estimators agree that sampling Optimization 34(3):441–466.
on the right-hand side of the domain is the most promising course of
action, but our formulation of noisy expected improvement prefers a less
explored region. This decision is motivated by the interesting behavior
of the posterior, which shows considerable disagreement regarding the
updated posterior mean; see figure 8.6. This nuance is only revealed
as our formulation reasons about the joint predictive distribution of y,0
whereas the plug-in estimators only inspect the marginals.
Another proposed approximation scheme for noisy expected im- approximating through reinterpolation
provement is reinterpolation. We fit a noiseless Gaussian process to im-
puted values of the objective function at the observed locations 𝝓 = 𝑓 (x),
then compute the exact expected improvement for this surrogate. A natu- 15 a. i. j. forrester et al. (2006). Design and
ral choice considered by forrester et al. is to impute using the posterior Analysis of “Noisy” Computer Experiments.
aiaa Journal 44(10):2331–2339.
mean:15
𝝓 ≈ 𝜇D (x),
resulting in the approximation (computed with respect to the surrogate):

𝛼 ei (𝑥; D) ≈ 𝛼 ei (𝑥; x, 𝝓). (8.19)

This procedure is illustrated in figure 8.5. The resulting decision is very

similar to that made by noisy expected improvement.
letham et al. also promoted this basic approach, but proposed marginal- 16 b. letham et al. (2019). Constrained Bayesian
izing rather than imputing the latent objective function values:16 Optimization with Noisy Experiments. Bayes-
ian Analysis 14(2):495–519.
∫
𝛼 ei (𝑥; D) ≈ 𝛼 ei (𝑥; x, 𝝓) 𝑝 (𝝓 | x, D) d𝝓. (8.20)

The approximate acquisition function is the expectation of the exact

expected improvement if we had access to exact observations. Although
this integral cannot be computed exactly, the authors described a straight-
forward and effective quasi-Monte Carlo approximation. letham et al.’s
approximation is illustrated in figure 8.6. There is good agreement with
the expected improvement acquisition function from figure 8.4, except
near the observation location chosen by that policy.
This stark disagreement is a consequence of reinterpolation: the reduction to zero at observed locations
acquisition function vanishes at previously observed locations – just
as the exact expected improvement does – regardless of the observed
17 d. huang et al. (2006b). Global Optimization
values. Thus repeated measurements at the same location are barred, of Stochastic Black-Box Systems via Sequen-
strongly encouraging exploration. huang et al. cite this property as tial Kriging Meta-Models. Journal of Global
undesirable,17 as repeated measurements can reinforce our understanding Optimization 34(3):441–466.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 165

 computing policies with gaussian processes

forrester et al.’s approximation (8.19) next observation location

Figure 8.5: forrester et al.’s approximation to noisy expected improvement (8.19). Given a Gaussian process fit to noisy
data, we compute the exact expected improvement (8.9) for a noiseless Gaussian process fit to the posterior
mean.

of the optimum by reducing uncertainty. letham et al. on the other hand

suggest the exploration boost is beneficial for optimization and point
out we may reduce uncertainty through measurements in neighboring
locations if desired.
adjusting for observation noise A weakness shared by all these approximations is that the underlying
noiseless expected improvement incorrectly assumes that our observa-
tion will reveal the exact objective value. In fact, observation noise is
ignored entirely, as all these acquisition functions are expectations with
respect to the unobservable quantity 𝜙 rather than the observed quantity
𝑦. This represents a disconnect between the reasoning of the optimiza-
18 d. huang et al. (2006b). Global Optimization tion policy and the true nature of the observation process. huang et al.
of Stochastic Black-Box Systems via Sequen- acknowledged this issue and proposed an augmented expected improve-
tial Kriging Meta-Models. Journal of Global
Optimization 34(3):441–466.
ment measure accounting for observation noise18 by multiplying by the
factor 1 − 𝜎𝑛 /𝑠,19 penalizing locations with low signal-to-noise ratios.
19 This is only sensible when the signal-to-noise
ratio is at least one. Ignoring the distribution of the observed value can be especially
problematic with heteroskedastic noise, as shown in figure 8.7. Both the
augmented18 and noisy expected improvement acquisition functions are
biased towards the left-hand side of the domain, where observations
reveal more information. Plug-in approximations, on the other hand, are
oblivious to this distinction and elect to explore the noisy region on the
right; see figure 8.4.
Our opinion is that all approximation schemes based on reducing to
exact expected improvement should be avoided with significant noise
levels, but can be a reasonable choice otherwise. With low noise, observed

166
 8.3. probability of improvement

sampled acquisition functions

letham et al.’s approximation (8.20) next observation location

Figure 8.6: letham et al.’s approximation to noisy expected improvement (8.20). We take the expectation of the exact
expected improvement (8.9) for a noiseless Gaussian process fit to exact observations at the observed locations.
The middle panels show realizations of the reinterpolated process and the resulting expected improvement.

values cannot stray too far from the true underlying objective function
value. Thus we have 𝑦 ≈ 𝜙; 𝑠 ≈ 𝜎, and any inaccuracy in (8.17), (8.18), or
(8.20) will be minor. However, this heuristic argument breaks down in
high-noise regimes.

8.3 probability of improvement

Probability of improvement (7.5) represents the probability that the sim- probability of improvement: § 7.5, p. 131
ple reward of our data (6.3) will exceed a threshold 𝜏 after obtaining an
observation at 𝑥:


𝛼 pi (𝑥; D) = Pr 𝑢 (D 0) > 𝜏 | 𝑥, D .

Like expected improvement, this quantity can be computed exactly for

our chosen model class.

Probability of improvement without noise

The probability of improvement after an exact observation at 𝑥 is simply
the probability that the observed value will exceed 𝜏 (7.6). For a Gaussian

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 167

computing policies with gaussian processes

augmented expected improvement (huang et al., 2006b)

noisy expected improvement (8.16)

Figure 8.7: A comparison of noiseless expected improvement with a plug-in estimate for 𝜙 ∗ (8.17) and the expected
one-step marginal gain in simple reward (7.2). Here the noise variance increases linearly from zero on the
left-hand side of the domain to a signal-to-noise ratio of 1 on the right-hand side. The larger enveloping area
shows 95% credible intervals for the noisy observation 𝑦, whereas the smaller area provides the same credible
intervals for the objective function.

process, this probability is given by the complementary Gaussian cdf:

𝜇 −𝜏 
𝛼 pi (𝑥; D, 𝜏) = Φ . (8.21)
𝜎
equivalent acquisition function, 𝛼 pi
0 As our policy will ultimately be determined by maximizing the probabil-
ity of improvement, it is prudent to transform this expression into the
simpler and better-behaved acquisition function
𝜇 −𝜏
𝛼 pi0 (𝑥; D, 𝜏) = , (8.22)
𝜎
which shares the same maxima but is slightly cheaper to compute and
does not suffer from a vanishing gradient for extreme values of 𝜏.
partial derivatives with respect to predictive We may gain some insight into the behavior of probability of im-
distribution parameters provement by computing the gradient of this alternative expression (8.22)
with respect to the parameters of the predictive distribution:
𝜕𝛼 pi0 1 𝜕𝛼 pi0 𝜏 −𝜇
= ; = .
𝜕𝜇 𝜎 𝜕𝜎 𝜎2
We observe that probability of improvement is monotonically increasing
with 𝜇, universally encouraging exploitation. Probability of improvement

168
 8.3. probability of improvement

𝑝 (𝑧) Figure 8.8: Probability of improvement with noise. The red

max(a + b𝑧) curve represents the updated simple reward given
𝜏 the 𝑧-score of the observed value. The probability
of improvement is the probability this value ex-
ceeds a threshold 𝜏, the area of the shaded region.
𝑧
ℓ 𝑢

is also increasing with 𝜎 when the target value is greater than the predic-
tive mean, or equivalently when the probability of improvement is less
than 1/2. Therefore probability of improvement also tends to encourage
exploration in this typical case. However, when the predictive mean is
greater than the improvement threshold, probability of improvement is,
perhaps surprisingly, decreasing with 𝜎, discouraging exploration in this
favorable regime. Probability of improvement favors a relatively safe
option returning a certain but modest improvement over a more-risky risk aversion of probability of improvement:
alternative offering a potentially larger but less-certain improvement, as § 7.5, p. 132
demonstrated in figure 7.8.
With the partial derivatives computed above, we may readily com- gradient of probability of improvement
pute the gradient of (8.22) with respect to the proposed evaluation loca- without noise
tion 𝑥 in the noiseless case (8.6):
 
𝜕𝛼 pi0 1 𝜕𝜇 0 𝜕𝜎
= − 𝛼 pi .
𝜕𝑥 𝜎 𝜕𝑥 𝜕𝑥

Probability of improvement with noise

As with expected improvement, computing probability of improvement
with noisy observations is slightly more complicated than with exact
observations, and for the same reason. A noisy observation can affect
the posterior mean at every previously observed location, and thus
improvement to the simple reward can occur at any point. However,
we can compute probability of improvement exactly by adapting the
techniques we used to compute noisy expected improvement (8.16).
As before, the key observation is that the updated posterior mean is
linear in the observed value 𝑦 (8.11), allowing us to express the updated
simple reward as
𝑢 (D 0) = max(a + b𝑧),
where a and b are defined in (8.12) and 𝑧 is the 𝑧-score of the observation.
Now we can write the probability of improvement as the expectation of
an indicator against a standard normal random variable; see figure 8.8:
∫
 
𝛼 pi (𝑥; D) = max(a + b𝑧) > 𝜏 𝜙 (𝑧) d𝑧.

Due to the convexity of the updated simple reward, improvement 20 In the case that improvement is impossible
occurs on at-most two intervals:20 or occurs at a single point, the probability of
improvement is zero.
(−∞, ℓ) ∪ (𝑢, ∞).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 169

 computing policies with gaussian processes

The endpoints of these intervals may be computed directly by inspecting

the intersections of the lines a + b𝑧 with the threshold:
 
ℓ = max (𝜏 − 𝑎𝑖 )/𝑏𝑖 | 𝑏𝑖 < 0 ; 𝑢 = min (𝜏 − 𝑎𝑖 )/𝑏𝑖 | 𝑏𝑖 > 0 ;
𝑖 𝑖

Note that one of these end points may not exist, for example if every
slope in the b vector were positive; see figure 8.3 for an example. In
this case we may take ℓ = −∞ or 𝑢 = ∞ as appropriate. Now given the
endpoints (ℓ, 𝑢), the probability of improvement may be computed in
terms of the standard normal cdf:
𝛼 pi (𝑥; D) = Φ(ℓ) + Φ(−𝑢). (8.23)
gradient of noisy probability of improvement: We may compute the gradient of this noisy formulation of probability
§ c.3, p. 309 of improvement; details are given in the appendix.

8.4 upper confidence bound

upper confidence bound: § 7.8, p. 145 Computing an upper confidence bound (7.18) for a Gaussian process is
trivial. Given a confidence parameter 𝜋 ∈ (0, 1), we must compute a
pointwise upper bound for the objective function with that confidence:
𝛼 ucb (𝑥; D, 𝜋) = 𝑞(𝜋; 𝑥, D), (8.24)
where 𝑞 is the quantile function of the predictive distribution. For a
Gaussian process, this quantile takes the simple form
𝛼 ucb (𝑥; D, 𝜋) = 𝜇 + 𝛽𝜎, (8.25)
exploration parameter, 𝛽 where 𝛽 = Φ−1 (𝜋) depends on the confidence level and can be computed
from the inverse Gaussian cdf. This acquisition function is normally
21 A word of warning: the√︁ exploration parameter parameterized directly in terms of 𝛽 rather than the confidence level 𝜋.
is sometimes denoted 𝛽, so that 𝛽 is a weight In this case 𝛽 can be interpreted as an “exploration parameter,” as higher
on the variance 𝜎 2 instead.
values clearly reward uncertainty more than smaller values.21 No special
care is required in computing (8.25), and its gradient with respect to the
gradient of upper confidence bound proposed observation location 𝑥 can also be computed easily:
𝜕𝛼 ucb 𝜕𝜇 𝜕𝜎
= +𝛽 .
𝜕𝑥 𝜕𝑥 𝜕𝑥

Correspondence with probability of improvement

22 d. r. jones (2001). A Taxonomy of Global Op- For Gaussian processes, we can derive an intimate correspondence be-
timization Methods Based on Response Sur- tween the (noiseless) probability of improvement and upper confidence
faces. Journal of Global Optimization 21(4):345–
383. [§ 6]
bound acquisition functions. Namely, a point maximizing probability of
improvement for a given target also maximizes some statistical upper
23 z. wang et al. (2016a). Optimization as Esti-
mation with Gaussian Processes in Bandit Set-
bound of the objective and vice versa, a fact that has been noted by
tings. aistats 2016. [lemma 3.1] several authors, including jones22 and, independently, wang et al.23
To establish this result, let an arbitrary exploration parameter 𝛽 be
given. Consider a point optimizing the upper confidence bound:
𝑥 ∈ arg max 𝜇 + 𝛽𝜎,
𝑥 0 ∈X

170
 8.5. approximate computation for one-step lookahead

and define equivalent improvement target, 𝜏 (𝛽)

𝜏 (𝛽) = max
0
𝜇 + 𝛽𝜎
𝑥 ∈X

to be its optimal value. Then the following equalities are satisfied at 𝑥:

𝜏 (𝛽) − 𝜇
𝜏 (𝛽) = 𝜇 + 𝛽𝜎; 𝛽= .
𝜎
Now it is easy to show that 𝑥 also minimizes the score
𝜏 (𝛽) − 𝜇
𝑥 ∈ arg min , (8.26)
𝑥 0 ∈X 𝜎

because if there were some other point 𝑥 0 with

𝜏 (𝛽) − 𝜇 0
< 𝛽,
𝜎0
then we would have
𝜇 0 + 𝛽𝜎 0 > 𝜏 (𝛽),
contradicting the optimality of 𝑥. Therefore 𝑥 also maximizes probability
of improvement with target 𝜏 (𝛽) (8.22).
It is important to note that the value of this target 𝜏 (𝛽) is data- and data-dependence of relationship
model-dependent and may change from iteration to iteration. There-
fore sequentially maximizing an upper confidence bound with a fixed
exploration parameter is not equivalent to maximizing probability of
improvement with a fixed improvement target.

8.5 approximate computation for one-step lookahead

Unfortunately we have exhausted the acquisition functions for which
exact computation is possible with Gaussian process models. However,
we can still proceed effectively with appropriate approximations. We will
begin by discussing the implementation of arbitrary one-step lookahead
policies when exact computation is not possible.
Recall that one-step lookahead entails maximizing the expected one-step lookahead: § 5.3, p. 101
marginal gain to a utility function 𝑢 (D) after making an observation at
a proposed location 𝑥:
∫
 
𝛼 (𝑥; D) = 𝑢 (D 0) − 𝑢 (D) N (𝑦; 𝜇, 𝑠 2 ) d𝑦.

If this integral is intractable, we must resort to analytic approxima-

tion or numerical integration to evaluate and optimize the acquisition
function. Fortunately in the case of sequential optimization, this is a
one-dimensional integral that can be approximated using standard tools.
It will be convenient below to introduce simple notation for the marginal gain from observation (𝑥, 𝑦),
marginal gain in utility resulting from a putative observation (𝑥, 𝑦). We Δ(𝑥, 𝑦)
will write
Δ(𝑥, 𝑦) = 𝑢 (D 0) − 𝑢 (D)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 171

 computing policies with gaussian processes

for this quantity, leaving the dependence on D implicit. Now we seek to

approximate ∫
𝛼 (𝑥; D) = Δ(𝑥, 𝑦) N (𝑦; 𝜇, 𝑠 2 ) d𝑦. (8.27)

Gauss–Hermite quadrature We recommend using off-the-shelf quadrature methods to estimate

the expected marginal gain (8.27).24 The most natural approach is Gauss–
24 p. j. davis and p. rabinowitz (1984). Methods Hermite quadrature, a classical approach for approximating integrals of
of Numerical Integration. Academic Press. the form ∫
𝐼 = ℎ(𝑧) exp(−𝑧 2 ) d𝑧. (8.28)

integration nodes, {𝑧𝑖 } Like all numerical integration methods, Gauss–Hermite quadrature en-
tails measuring the integrand ℎ at a set of 𝑛 points, called nodes, {𝑧𝑖 },
then approximating the integral by a weighted sum of the measured
𝑛=5 values:
𝑛=8
∑︁
𝑛
𝐼≈ 𝑤𝑖 ℎ(𝑧𝑖 ).
𝑖=1

Remarkably, this estimator is exact when the integrand is a polynomial

𝑧 of degree less than 2𝑛 − 1. This guarantee provides some guidance for
Gauss–Hermite nodes/weights for 𝑛 ∈ {5, 8}. selecting the order of the quadrature rule depending on how well the
marginal gain Δ may be approximated by a polynomial over the range
25 t. s. shao et al. (1964). Tables of Zeros and of plausible observations. Tables of integration nodes and quadrature
Gaussian Weights of Certain Associated La- weights for 𝑛 up to 64 are readily available, although such a high order
guerre Polynomials and the Related Gener-
alized Hermite Polynomials. Mathematics of
should rarely be needed.25
Computation 18(88):598–616. Through a simple transformation,26 we can rewrite the arbitrary
26 We take Gaussian expectation (8.27) in the Gauss–Hermite form (8.28):
√ 𝑦−𝜇 ∫ ∫
𝑦 = 𝜇 + 2𝑠𝑧 ⇔ 𝑧= √ ; 1 √
2𝑠 Δ(𝑥, 𝑦) N (𝑦; 𝜇, 𝑠 2 ) d𝑦 = √ Δ(𝑥, 𝜇 + 2𝑠𝑧) exp(−𝑧 2 ) d𝑧.
𝜋
Note we√must account for the normalization
factor ( 2𝜋𝜎) −1 of the Gaussian distribution,
hence the constant that appears.
If we define appropriately renormalized weights
√
𝑤¯ 𝑖 = 𝑤𝑖 / 𝜋,

then we arrive at the following approximation to the acquisition function:

∑︁
𝑛 √
𝛼 (𝑥; D) ≈ 𝑤¯ 𝑖 Δ(𝑥, 𝑦𝑖 ); 𝑦𝑖 = 𝜇 + 2𝑧𝑖 𝑠. (8.29)
𝑖=1

approximating gradient via Gauss–Hermite We may also extend this scheme to approximate the gradient of the acqui-
quadrature: § c.3, p. 309 sition function using the same nodes and weights; details are provided
in the accompanying appendix.

8.6 knowledge gradient

knowledge gradient: § 7.4, p. 129 The knowledge gradient is the expected one-step gain in the global
reward (6.5). If we define

𝜇 ∗ = max 𝜇D (𝑥)
𝑥 ∈X

172
 8.6. knowledge gradient

to be the utility of the current dataset, then we must compute:

∫h i
𝛼 kg (𝑥; D) = max
0
𝜇 D
0
0 (𝑥 ) − 𝜇
∗
𝑝 (𝑦 | 𝑥, D) d𝑦. (8.30)
𝑥 ∈X

The global optimization in the expectation renders the knowledge gradi-

ent nontrivial to compute in most cases, so we must resort to quadrature
or other approximation.

Exact computation in discrete domains

frazier et al. first proposed the knowledge gradient for optimization on 27 p. frazier et al. (2009). The Knowledge-
a discrete domain X = {1, 2, . . . , 𝑛}.27 In this case, the objective is simply Gradient Policy for Correlated Normal Beliefs.
informs Journal on Computing 21(4):599–613.
a vector f ∈ ℝ𝑛, and a Gaussian process belief about the objective is
simply a multivariate normal distribution:

𝑝 (f | D) = N (f; 𝝁, 𝚺).

The knowledge gradient now reduces to the expected marginal gain in

the maximum of the posterior mean vector:
∫
 
𝛼 kg (𝑥; D) = max 𝝁 0 − 𝜇 ∗ 𝑝 (𝑦 | 𝑥, D) d𝑦.

We may compute this expectation in closed form following our analysis computation of noisy expected improvement:
of noisy expected improvement, which was merely a slight adaptation § 8.2, p. 160
of frazier et al.’s approach.
The updated posterior mean vector 𝝁 0 after observing an observation
(𝑥, 𝑦) is linear in the observed value 𝑦:
𝚺𝑥 𝑦 − 𝜇𝑥
𝝁0 = 𝝁 + ,
𝑠 𝑠
where 𝜇𝑥 is the entry of 𝝁 corresponding to the index 𝑥, and 𝚺𝑥 is
similarly the corresponding column of 𝚺. If we define
𝚺𝑥
a = 𝝁; b= ,
𝑠
then we may rewrite the knowledge gradient in terms of the 𝑔 function
introduced in the context of noisy expected improvement (8.10): computation in terms of 𝑔 (a, b)

𝛼 kg (𝑥; D) = 𝑔(a, b) − 𝜇 ∗.

We may evaluate this expression exactly in O(𝑛 log 𝑛) time following computation of 𝑔: § 8.2, p. 161
our previous discussion. Thus we may compute the knowledge gradi-
ent policy with a discrete domain in O(𝑛 2 log 𝑛) time per iteration by
exhaustive computation of the acquisition function.

Approximation via numerical quadrature

Although the knowledge gradient acquisition function can be computed
exactly over discrete domains, the situation becomes significantly more

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 173

 computing policies with gaussian processes

𝑥 𝑥0
𝑝 (𝑧)
𝑥 𝑥0

Figure 8.9: The complex behavior of the updated global re-

ward. Above: the location of the posterior mean 𝑝 (𝑧)
maximum given the 𝑧-score of an observation
Δ(𝑥, 𝑧)
at two points, 𝑥 and 𝑥 .0 An observation at 𝑥 0 al-
Δ(𝑥 ,0 𝑧)
ways results in shoring up the existing maximum,
whereas an observation at 𝑥 reveals a new maxi-
mum given a sufficiently high observation. Below:
𝑧
the marginal gain in the global reward as a function
of the 𝑧-score of an observation at these locations.

complicated in continuous domains. The culprit is the nonconvex global

optimization in (8.30), which makes the knowledge gradient intractable
except in a few special cases.
exact computation for special cases Some Gaussian processes give rise to convenient structure in the
posterior distribution facilitating computation of the knowledge gradi-
ent. For example, in one dimension the Wiener or Ohrstein–Uhlenbeck
processes satisfy a Markov property guaranteeing the posterior mean is
always maximized at an observed location. As a result the simple reward
28 h. j. kushner (1964). A New Method of Locat- and global reward are always equal, and the knowledge gradient reduces
ing the Maximum Point of an Arbitrary Multi- to expected improvement. This structure was often exploited in the early
peak Curve in the Presence of Noise. Journal
of Basic Engineering 86(1):97–106.
literature on Bayesian optimization.28, 29
Figure 8.9 give some insight into the complexity of the knowledge
29 j. mockus (1972). Bayesian Methods of Search
for an Extremum. Avtomatika i Vychisli- gradient integral (8.30). We illustrate the possible results from adding an
tel’naya Tekhnika (Automatic Control and Com- observation at two locations as a function of the 𝑧-score of the observed
puter Sciences) 6(3):53–62. value. For the point on the left, the behavior is similar to expected im-
provement: a sufficiently high value moves the maximum of the posterior
mean to that point; otherwise, we retain the incumbent. The marginal
gain in utility is a piecewise linear function corresponding to these two
outcomes, as in expected improvement. The point on the right displays
entirely different behavior – the marginal gain in utility is smooth, and
nearly any outcome would be beneficial.

174
 8.6. knowledge gradient

knowledge gradient, 𝛼 kg next observation location

kgcp approximation

Figure 8.10: A comparison of the knowledge gradient acquisition function and the kgcp approximation for an example
scenario. In this case the kgcp approximation reverts to expected improvement.

When exact computation is not possible, we must result to numerical

integration or an analytic approximation when adopting the knowledge
gradient policy. The former path was explored in depth by wu et al.30 To 30 j. wu et al. (2017). Bayesian Optimization with
compute the acquisition function at a given point, we can compute a high- Gradients. neurips 2017.
accuracy approximation following the numerical techniques outlined in
the previous section. Order-𝑛 Gauss–Hermite quadrature for example,
would approximate with (8.29):

1 ∑︁
𝑛
𝛼 kg (𝑥; D) ≈ √ 𝑤𝑖 Δ(𝑥, 𝑦𝑖 ); Δ(𝑥, 𝑦𝑖 ) = max 𝜇D0 (𝑥 0) − 𝜇 ∗.
𝜋 𝑖=1 0
𝑥 ∈X

With some care, we can also approximate the gradient of the knowl- approximating gradient of knowledge
edge gradient in this scheme; details are given in an appendix. gradient: § c.3, p. 310

Knowledge gradient for continuous parameters approximation

scott et al. suggested an alternative and lightweight approximation 31 w. scott et al. (2011). The Correlated Knowl-
scheme for the knowledge gradient the authors called the knowledge edge Gradient for Simulation Optimization of
Continuous Parameters Using Gaussian Pro-
gradient for continuous parameters (kgcp).31 The kgcp approximation cess Regression. siam Journal on Optimization
entails replacing the domain of the maximization in the definition of the 21(3):996–1026.
global reward utility, normally the entire domain X, with a conveniently
chosen discrete set: the already observed locations x and the proposed
new observation location 𝑥. Let x0 represent this set. We approximate
the current and future utility with

𝑢 (D) ≈ max 𝜇D (x0); 𝑢 (D 0) ≈ max 𝜇D0 (x0),

yielding the approximation

 
𝛼 kg (𝑥; D) ≈ 𝔼 max 𝜇D0 (x0) | 𝑥, D − max 𝜇D (x0). (8.31)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 175

 computing policies with gaussian processes

comparison with expected improvement This expression is almost identical to expected improvement (8.7)!
In fact, if we define
𝜇 ∗ = max 𝜇D (x),
then a simple manipulation gives:

𝛼 kg (𝑥; D) ≈ 𝛼 ei (𝑥; D) − max(𝜇 − 𝜇,∗ 0).

Effectively, the kgcp approximation is a simple adjustment to expected

improvement where we punish points for already having large expected
value. From the point of view of the global reward, these points already
represent success and their large expected values are already reflected in
the current utility. Therefore, we should not necessarily waste precious
evaluations confirming what we already believe.
gradient of kgcp approximation The gradient of the kgcp approximation may also be computed in
terms of the gradient of expected improvement and the posterior mean:
𝜕𝛼 kg 𝜕𝛼 ei 𝜕𝜇
≈ − [𝜇 > 𝜇 ∗ ] .
𝜕𝑥 𝜕𝑥 𝜕𝑥
example and discussion In the case of our example scenario, the posterior mean never ex-
ceeds the highest observed point. Therefore, the kgcp approximation to
the knowledge gradient globally reduces to the expected improvement;
compare with the expected improvement in figure 7.3. Comparing with
true knowledge gradient, we can see that this approximation cannot
necessarily be trusted to be unconditionally faithful. However, the kgcp
approximation has the advantage of efficient computation and may be
used as a drop-in replacement for expected improvement that may offer
a slight boost in performance when the global reward utility is preferred
to simple reward.

8.7 thompson sampling

Thompson sampling: § 7.9, p. 148 Thompson sampling designs each observation by sampling a point pro-
portional to its probability of maximizing the objective (7.19):

𝑥 ∼ 𝑝 (𝑥 ∗ | D).

A major barrier to Thompson sampling with Gaussian processes is the

complex nature of this distribution. Except in a small number of special
32 One notable example is the Wiener process, cases,32 this distribution cannot be computed analytically. Figure 8.11
where 𝑥 ∗ famously has an arcsine distribution: illustrates the complicated nature of this distribution for our running
p. lévy (1948). Processus stochastiques et mou- example, which was only revealed via brute-force sampling. However, a
vement brownien. Gauthier–Villars. straightforward implementation strategy is to maximize a draw from the
objective function posterior, which assumes the role of an acquisition
function:
𝛼 ts (𝑥; D) ∼ 𝑝 (𝑓 | D).
The global optimum of 𝛼 ts is then a sample from the desired distribution.
This procedure in fact yields a sample from the joint distribution
of the location and value of the optimum, 𝑝 (𝑥 ,∗ 𝑓 ∗ | D), as the value of

176
 8.7. thompson sampling

𝑝 (𝑥 ∗ | D)

Thompson samples

Figure 8.11: The distribution of the location of the global maximum, 𝑝 (𝑥 ∗ | D), for an example scenario, and 100 samples
drawn from this distribution.

the sampled objective function 𝛼 ts at its maximum provides a sample 𝑓∗

from 𝑝 (𝑓 ∗ | 𝑥 ,∗ D); see the margin. We discuss Thompson sampling now
because the ability to sample from these distributions will be critical for
computing mutual information, our focus in the following sections.

Exhaustive sampling 𝑥∗
In “small” domains, we can realize Thompson sampling via brute force. Maximizing a draw from a Gaussian process
If the domain can be exhaustively covered by a sufficiently small set of naturally samples from 𝑝 (𝑥 ,∗ 𝑓 ∗ | D)
points 𝝃 (for example, with a dense grid or a low-discrepancy sequence)
then we can simply sample the associated objective function values
𝝓 = 𝑓 (𝝃 ) and maximize:33 33 We are taking a slight liberty with notation
here as we have previously used 𝝓 for 𝑓 (x),
𝑥 = arg max 𝝓; 𝝓 ∼ 𝑝 (𝝓 | 𝝃 , D). the latent objective function values at the ob-
served locations. However, for the remainder
The distribution of 𝝓 is multivariate normal, making sampling easy: of the we will be assuming the general, poten-
tially noisy case where our data will be written
D = (x, y) and we will have no need to refer
𝑝 (𝝓 | 𝝃 , D) = N (𝝓; 𝝁, 𝚺); 𝝁 = 𝜇D (𝝃 ); 𝚺 = 𝐾D (𝝃 , 𝝃 ). to 𝑓 (x).

The running time of this procedure grows quickly with the size of 𝝃 , 34 g. pleiss et al. (2020). Fast Matrix Square Roots
although sophisticated numerical methods enable scaling to roughly with Applications to Gaussian Processes and
Bayesian Optimization. neurips 2020.
50 000 points.34
Figure 8.11 shows 100 Thompson samples for our example scenario
generated via exhaustive sampling, taking 𝝃 to be a grid of 1000 points.

On-demand sampling
An alternative to exhaustive sampling is to use off-the-shelf optimization
routines to maximize a draw from the objective function posterior we

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 177

 computing policies with gaussian processes

Algorithm 8.1: On-demand sampling. Dts ← D I initialize fictitious dataset with current data
repeat
given request for observation at 𝑥:
𝜙 ← 𝑝 (𝜙 | 𝑥, Dts ) I sample value at 𝑥
Dts ← Dts ∪ (𝑥, 𝜙) I update fictitious dataset
yield 𝜙
until external optimizer terminates

build progressively on demand. Namely, when an optimization routine

requests an evaluation of 𝛼 ts at a point 𝑥, we sample an objective function
value at that location:
𝜙 ∼ 𝑝 (𝜙 | 𝑥, D),
then augment our dataset with the simulated observation (𝑥, 𝜙). We
proceed in this manner until the optimizer terminates. This procedure
avoids simulating the entire objective function while guaranteeing the
joint distribution of the provided evaluations is correct via the chain
rule of probability. Pseudocode is provided in Algorithm 8.1 in the form
of a generator function; the “yield” statement returns a value while
low-rank updates: § 9.1, p. 202 maintaining state. Using rank-one updates to update the posterior, the
running time computational cost of generating each evaluation scales quadratically
with the size of the fictitious dataset Dts .
using gradient methods If desired, we may use gradient methods to optimize the generated
sample by sampling from the joint posterior of the function value and
its gradient at each requested location:

𝑝 (𝜙, ∇𝜙 | 𝑥, Dts )

35 m. lázaro-gredilla et al. (2010). Sparse Spec- However, in high dimensions, the additional cost required to condition
trum Gaussian Process Regression. Journal of on these gradient observations may become excessive. In 𝑑 dimensions,
Machine Learning Research 11(Jun):1865–1881.
returning gradients effectively reduces the number of function evalua-
36 j. m. hernández-lobato et al. (2014). Predic- tions we can allow for the optimizer by a factor of (𝑑 + 1) if we wish to
tive Entropy Search for Efficient Global Opti-
maintain the same total computational effort.
mization of Black-box Functions. neurips 2014.

Sparse spectrum approximation for stationary covariance functions

stationarity: § 3.2, p. 50 If the prior covariance function 𝐾 of the Gaussian process is stationary,
we may use a sparse spectrum approximation35 to the posterior Gaussian
process to dramatically accelerate Thompson sampling.36
Consider a stationary covariance function 𝐾 on ℝ𝑑 with spectral
37 Recall the convention of writing a station- density 𝜅. The key idea in sparse spectrum approximation is to interpret
ary covariance with respect to a single input, the characterization in (3.10) as an expectation with respect to the spectral
𝐾 (x, x0 ) = 𝐾 (x − x0 ).
density:37


𝐾 (x − x0) = 𝐾 (0) 𝔼𝝃 exp 2𝜋𝑖 (x − x0)>𝝃 , (8.32)

and approximate via Monte Carlo integration. We first sample a set of 𝑚

spectral points, {𝝃𝑖 } frequencies, called spectral points, from the spectral density: {𝝃𝑖 } ∼ 𝜅 (𝝃 ).

178
 8.7. thompson sampling

posterior mean, exact

posterior 95% credible interval, exact
Figure 8.12: Sparse spectrum approxima-
tion. Top: the exact posterior
belief (about noisy observa-
tions rather than the latent
function) for a Gaussian pro-
cess conditioned on 200 ob-
servations. Bottom: a sparse
posterior mean, sparse spectrum approximation spectrum approximation us-
posterior 95% credible interval, sparse spectrum approximation ing 100 spectral points sam-
pled from the spectral den-
sity.

To enforce the symmetry around the origin inherent to the spectral

density (theorem 3.2), we augment each sample 𝝃𝑖 with its negation, −𝝃𝑖 .
Using these samples for a Monte Carlo approximation to (8.32) has
the effect of approximating a Gaussian process GP (𝑓 ; 𝜇, 𝐾) with a finite-
dimensional Gaussian process:

𝑓 (x) ≈ 𝜇 (x) + 𝜷 >𝝍 (x). (8.33)

Here 𝜷 is a vector of normally distributed weights and 𝝍 : ℝ𝑑 → ℝ2𝑚 is weight vector, 𝜷

a feature representation determined by the spectral points: feature representation, 𝝍



𝜓 2𝑖−1 (x) = cos 2𝜋 𝝃𝑖> x ; 𝜓 2𝑖 (x) = sin 2𝜋 𝝃𝑖> x .

For the additive Gaussian noise model, if N is the covariance of the noise noise covariance matrix, N
contributions to the observed values y, the posterior moments of the
weight vector in this approximation are:

𝔼[𝜷 | D] = 𝚿> (𝚿𝚿> + 𝑎 −2 N) −1 (y − 𝝁);

 
cov[𝜷 | D] = 𝑎 2 I − 𝚿> (𝚿𝚿> + 𝑎 −2 N) −1 𝚿 ,

where 𝑎 2 = 𝐾 (0)/𝑚, 𝝁 is the prior mean of y, and 𝚿 is a matrix whose

rows comprise the feature representations of the observed locations. training features, 𝚿
The sparse spectrum approximation allows us to generate a posterior
sample of the objective function in time O(𝑛𝑚 2 ) by drawing a sample sampling from a multivariate normal
from the weight posterior and appealing to the representation in (8.33). distribution: § a.2, p. 299
The resulting sample is nontrivial to maximize due to the nonlinear
nature of the representation, but with the sampled weights in hand, we
can generate requested function and gradient evaluations in constant
time, enabling exhaustive optimization via gradient methods.
A sparse spectrum approximation is illustrated in figure 8.12, using a
total of 100 spectral points. The approximation is quite reasonable near
data and acceptable in extrapolatory regions as well.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 179

 computing policies with gaussian processes

8.8 mutual information with 𝑥 ∗

mutual information: § 7.6, p. 135 Mutual information measures the expected information gain (6.3) pro-
vided by an observation at 𝑥 about some random variable of interest 𝜔,
which can be expressed in two equivalent forms (7.14–7.15):
 
𝛼 mi (𝑥; D) = 𝐻 [𝜔 | D] − 𝔼𝑦 𝐻 [𝜔 | D 0] | 𝑥, D (8.34)
 
= 𝐻 [𝑦 | 𝑥, D] − 𝔼𝜔 𝐻 [𝑦 | 𝜔, 𝑥, D] | 𝑥, D . (8.35)
In the context of Bayesian optimization, the most natural choices for 𝜔
are the location 𝑥 ∗ and value 𝑓 ∗ of the global optimum, both of which
were discussed in the previous chapter. Unfortunately, a Gaussian process
belief on the objective function in general induces complex distributions
for these quantities, which makes even approximating mutual infor-
mation somewhat involved. However, several effective approximation
schemes are available for both options. We will consider each in turn,
beginning with 𝑥 ∗.

Direct form of mutual information

Of the two equivalent formulations of mutual information above, the
former (8.34) is perhaps the most natural, as it reasons directly about
38 j. villemonteix et al. (2009). An informa- changes in the entropy of the variable of interest. Initial work on mutual
tional approach to the global optimization information with 𝑥 ∗ considered approximations to this direct expres-
of expensive-to-evaluate functions. Journal of
Global Optimization 44(4):509–534.
sion38 – including the work coining the now-common moniker entropy
search for information-theoretic optimization policies39 – but this ap-
39 p. hennig and c. j. schuler (2012). Entropy
Search for Information-Efficient Global Op-
proach has fallen out of favor in preference for the latter formulation
timization. Journal of Machine Learning Re- (8.35), discussed below.
search 13(Jun):1809–1837. The main computational difficulty in approximating (8.34) is the un-
wieldy (and potentially high-dimensional) distribution 𝑝 (𝑥 ∗ | D); see
even the simple example in figure 8.11. There is no general method for
computing the entropy of this distribution in closed form, and overcom-
ing this barrier requires complex approximation schemes. The usual
approach is to make a discrete approximation to this distribution via
a set of carefully maintained so-called representer points, then reason
about changes in the entropy of this surrogate distribution via further
layers of approximation.

Predictive form of mutual information

40 j. m. hernández-lobato et al. (2014). Predic- hernández-lobato et al. proposed a faithful approximation to the mu-
tive Entropy Search for Efficient Global Opti- tual information with 𝑥 ∗ based on the formulation in (8.35):40
mization of Black-box Functions. neurips 2014.
 
𝛼𝑥 ∗ (𝑥; D) = 𝐻 [𝑦 | 𝑥, D] − 𝔼 𝐻 [𝑦 | 𝑥, 𝑥 ,∗ D] | 𝑥, D (8.36)
Compared to the direct form (8.34), this formulation is attractive as
all entropy computations are restricted to the one-dimensional predic-
tive distribution 𝑝 (𝑦 | 𝑥, D) rather than the inconvenient distribution
predictive entropy search 𝑝 (𝑥 ∗ | D). The authors named their approach predictive entropy search
to highlight this feature.

180
 8.8. mutual information with 𝑥 ∗

Let us consider the evaluation of (8.36) for a Gaussian process model computing first term
with additive Gaussian noise. To begin, the first term is simply the dif-
ferential entropy of a one-dimensional Gaussian distribution (8.5) and
may be computed in closed form (a.17):

𝐻 [𝑦 | 𝑥, D] = 1
2 log(2𝜋𝑒𝑠 2 ).

Unfortunately, the second term is significantly more complicated to work

with, and we can identify two primary challenges.

Approximating expectation with respect to 𝑥 ∗

First we must compute an expectation with respect to the location of the
global optimum 𝑥 ∗. The complexity of its distribution limits our options,
but Monte Carlo integration remains a viable option. We use Thompson Thompson sampling for gps: § 8.7, p. 176
sampling to generate samples of the optimal location {𝑥𝑖∗ }𝑛𝑖=1 ∼ 𝑝 (𝑥 ∗ | D)
and estimate the expected updated entropy with:

  1 ∑︁𝑛
𝔼 𝐻 [𝑦 | 𝑥, 𝑥 ,∗ D] | 𝑥, D ≈ 𝐻 [𝑦 | 𝑥, 𝑥𝑖∗, D].
𝑛 𝑖=1

When the prior covariance function is stationary, hernández-lobato sparse spectrum approximation for
et al. further propose to exploit the efficient approximate Thompson Thompson sampling: § 8.7, p. 178
sampling scheme via sparse spectrum approximation described in the
last section. When feasible, this reduces the cost of drawing the samples,
but it is not necessary for correctness.

Gaussian approximation to conditional predictive distribution

Now we must address the predictive distribution conditioned on the
location of the global optimum, 𝑝 (𝑦 | 𝑥, 𝑥 ,∗ D), which we may express as
the result of marginalizing the latent objective value 𝜙:
∫
𝑝 (𝑦 | 𝑥, 𝑥 ,∗ D) = 𝑝 (𝑦 | 𝑥, 𝜙) 𝑝 (𝜙 | 𝑥, 𝑥 ,∗ D) d𝜙 . (8.37)

It is unclear how we can condition our belief on the objective function

given knowledge of the optimum, and – as the resulting posterior on 𝜙
will be non-Gaussian – how we can resolve the resulting integral (8.37).
A key insight is that if the predictive distribution 𝑝 (𝜙 | 𝑥, 𝑥 ,∗ D) 41 In this case, the integral in (8.37) becomes the
were Gaussian, we could compute (8.37) in closed form,41 suggesting a convolution of two Gaussians, which may be
interpreted as the distribution of the sum of
promising path forward. In particular, consider an arbitrary Gaussian independent Gaussian random variables – see
approximation: § a.2, p. 300.
𝑝 (𝜙 | 𝑥, 𝑥 ,∗ D) ≈ N (𝜙; 𝜇∗, 𝜎∗2 ), (8.38)
whose parameters depend 𝑥 ∗ as their subscripts indicate. Plugging into
(8.37), we may estimate the predictive variance of 𝑦 given 𝑥 ∗ with (a.15): approximate variance of 𝑦 given 𝑥 ,∗ 𝑠 ∗2

var[𝑦 | 𝑥, 𝑥 ,∗ D] ≈ 𝜎∗2 + 𝜎𝑛2 = 𝑠 ∗2, (8.39)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 181

 computing policies with gaussian processes

𝑥∗

Figure 8.13: The example scenario we will consider for illustrating the predictive entropy search approximation to
𝑝 (𝑓 | 𝑥, 𝑥 ,∗ D), using the marked location for 𝑥 ∗.

and thus its differential entropy with (a.17):

𝐻 [𝑦 | 𝑥, 𝑥 ,∗ D] ≈ 1
2 log(2𝜋𝑒𝑠 ∗2 ). (8.40)

After simplification, the resulting approximation to (8.36) becomes:

1 ∑︁
𝑛
𝛼𝑥 ∗ (𝑥; D) ≈ 𝛼 pes (𝑥; D) = log 𝑠 − log 𝑠 ∗𝑖 . (8.41)
𝑛 𝑖=1

Approximation via Gaussian expectation propagation

To realize a complete algorithm, we need some way of finding a suitable
Gaussian approximation to 𝑝 (𝜙 | 𝑥, 𝑥 ,∗ D), and hernández-lobato et al.
describe one effective approach. The high-level idea is to begin with
the objective function posterior 𝑝 (𝑓 | D), impose a series of constraints
implied by knowledge of 𝑥 ,∗ then approximate the desired posterior via
approximate inference. We will describe the procedure for an arbitrary
putative optimum 𝑥 ,∗ illustrating each step of the approximation for the
example scenario and assumed optimum location shown in figure 8.13.
For the moment, we will proceed with complete disregard for computa-
tional efficiency, and return to the question of implementation shortly.

Ensuring 𝑥 ∗ is a local optimum

We first condition our belief on 𝑥 ∗ being a local optimum by insisting
gradient, Hessian at 𝑥 ∗ : ∇,∗ H∗ that the point satisfy the second partial derivative test. Let ∇∗ and H∗
respectively represent the gradient and Hessian of the objective function
42 We discussed a simpler, one-dimensional ana- at 𝑥 ∗. Local optimality implies that the gradient is zero and the Hessian
log of this task in § 2.8, p. 39. is negative definite:42

∇∗ = 0; (8.42)
H∗ ≺ 0. (8.43)

gradient constraint: ∇∗ = 0 Enforcing the gradient constraint (8.42) is straightforward, as we can

directly condition on a gradient observation. We show the result for our
example scenario in the top panel of figure 8.14.
first Hessian constraint: diagonal entries are The Hessian constraint (8.43) however is nonlinear and we must
negative resort to approximate inference. hernández-lobato et al. approximate

182
 8.8. mutual information with 𝑥 ∗

𝑥∗

𝑥∗

Figure 8.14: Top: the posterior for our example after conditioning on the derivative being zero at 𝑥 ∗ (8.42). Bottom: the
approximate posterior after conditioning on the second derivative being negative at 𝑥 ∗ (8.44).

this condition by breaking it into two complimentary components. First

we compel every diagonal entry of the Hessian to be negative. Letting
h∗ = diag H∗ ; we assume: diagonal of Hessian at 𝑥 ,∗ h∗

∀𝑖 : ℎ𝑖∗ < 0. (8.44)

We then fix the off-diagonal entries of the Hessian to values of our

choosing. One simple option is to set all off-diagonal entries to zero: second Hessian constraint: off-diagonal
entries are fixed
upper H∗ = 0, (8.45)

which combined with the diagonal constraint guarantees negative def- 43 hernández-lobato et al. point out this may
initeness.43 The combination of these conditions (8.44–8.45) is stricter not be faithful to the model and suggest the
alternative of matching the off-diagonal en-
than mere negative definiteness, as we eliminate all degrees of freedom tries of the Hessian of the objective function
for the off-diagonal entries. However, an advantage of this approach is sample that generated 𝑥 ∗. However, this does
that we can enforce the off-diagonal constraint via exact conditioning. not guarantee negative definiteness without
To proceed we dispense with the constraints we can condition on tweaking (8.44).
exactly. Let D 0 represent our dataset augmented with the gradient (8.42)
and off-diagonal Hessian (8.45) observations. The joint distribution of
the latent objective value 𝜙, the purportedly optimal value 𝜙 ∗ = 𝑓 (𝑥 ∗ ),
and the diagonal of the Hessian h∗ given this additional information is
multivariate normal:

𝑝 (𝜙, 𝜙 ,∗ h∗ | 𝑥 ,∗ D 0) = N (𝜙, 𝜙 ,∗ h∗; 𝝁 ,∗ 𝚺∗ ). (8.46)

We will now subject this initial belief to a series of factors corresponding

to desired nonlinear constraints. These factors will be compatible with
Gaussian expectation propagation, which we will use to finally derive Gaussian ep: § b.2, p. 302
the desired Gaussian approximation to the posterior (8.38). To begin, the truncating a variable with ep: § b.2, p. 305
Hessian diagonal constraint (8.44) contributes one factor for each entry:
Ö
[ℎ𝑖∗ < 0]. (8.47)
𝑖

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 183

 computing policies with gaussian processes

𝑥∗

𝑥∗

Figure 8.15: Top: the approximate posterior after conditioning on 𝜙 ∗ exceeding the function values at previously measured
locations (8.49). Bottom: the approximate posterior after conditioning on 𝜙 ∗ dominating elsewhere (8.50).

44 For this and the following demonstrations, we Our approximate posterior after incorporating (8.47) and performing
show the expectation propagation approxi- expectation propagation is shown in the bottom panel of 8.14.44
mation to the entire objective function pos-
terior; this is not required to approximate the
marginal predictive distribution and is only
for illustration.
Ensuring 𝑥 ∗ is a global optimum
Our belief now reflects our desire that 𝑥 ∗ be a local maximum; however,
we wish for 𝑥 ∗ to be the global maximum. Global optimality is not
easy to enforce, as it entails infinitely many constraints bounding the
objective at every point in the domain. hernández-lobato et al. instead
approximate this condition with optimality at the most relevant locations:
the already-observed points x and the proposed point 𝑥.
ensuring 𝜙 ∗ > max 𝝓 To enforce that 𝜙 ∗ exceed the objective function values at the ob-
served points, we could theoretically add 𝝓 = 𝑓 (x) to our prior (8.46),
Î
then add one factor for each observation: 𝑗 [𝜙 𝑗 < 𝜙 ∗ ]. However, this
approach requires an increasing number of factors as we gather more
data, rendering expectation propagation (and thus the acquisition func-
tion) increasingly expensive. Further, factors corresponding to obviously
suboptimal observations are uninformative and simply represent extra
work for no benefit.
truncating with respect to an unknown Instead, we enforce this constraint through a single factor truncating
threshold with ep: § b.2, p. 306 with respect to the maximal value of 𝝓: [𝜙 ∗ < max 𝝓]. In general, this
threshold will be a random variable unless our observations are noise-
less. Fortunately, expectation propagation enables tractable approximate
truncation at an unknown, Gaussian-distributed threshold. Define
2
𝜇max = 𝔼[max 𝝓 | D]; 𝜎max = var[max 𝝓 | D]; (8.48)

45 c. e. clark (1961). The Greatest of a Finite Set these moments can be approximated via either sampling or an assumed
of Random Variables. Operations Research 9(2): density filtering approach described by clark.45 Taking a moment-
145–162.
matched Gaussian approximation to max 𝝓 and integrating yields the

184
 8.8. mutual information with 𝑥 ∗

mutual information with 𝑥 ∗, 𝛼𝑥 ∗ next observation location

predictive entropy search approximation (8.41)

Figure 8.16: The predictive entropy search approximation (8.41) to the mutual information with 𝑥 ∗ acquisition function
(8.36) using the 100 Thompson samples from figure 8.11.

factor (b.13):46 46 With high noise, this approximation could be

𝜙∗ − 𝜇  improved slightly by computing moments of
max
Φ . (8.49) max 𝝓 given D,0 at additional expense.
𝜎max
We show the approximate posterior for our running example after incor-
porating this factor in the top panel of figure 8.15. The probability mass
at 𝑥 ∗ has shifted up dramatically.
Finally, to constrain 𝜙 ∗ to dominate the objective function value at a ensuring 𝜙 ∗ > 𝜙
point of interest 𝑥, we add one additional factor: enforcing order with ep: § b.2, p. 306

[𝜙 < 𝜙 ∗ ]. (8.50)

We obtain the final approximation to 𝑝 (𝜙 | 𝑥, 𝑥 ,∗ D) by combining the

prior in (8.46) with the factors (8.47, 8.49–8.50):
𝜙∗ − 𝜇 Ö
max
N (𝜙, 𝜙 ,∗ h∗ ; 𝝁 ,∗ 𝚺∗ ) [𝜙 < 𝜙 ∗ ] Φ [ℎ𝑖∗ < 0],
𝜎max 𝑖

approximating with Gaussian expectation propagation, and deriving

the marginal belief about 𝜙. The resulting final approximate posterior
for our example scenario is shown in the bottom panel of figure 8.15.
Our predictive uncertainty has been moderated in response to the final
constraint (8.50).
With the ability to approximate the latent predictive posterior, we
can now compute the predictive entropy search acquisition function
(8.41) via Thompson sampling and following the above procedure for
each sample. Figure 8.16 shows the approximation computed with 1000
Thompson samples. The approximation is excellent and induces a near-
optimal decision. Any deviation from the truth mostly reflects bias in
the Thompson sample distribution.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 185

 computing policies with gaussian processes

Efficient implementation
A practical realization of predictive entropy search can benefit from
careful precomputation and reuse of partial results. We will outline an
efficient implementation strategy, beginning with three steps of one-time
initial work.
1. Estimate the moments of max 𝝓 (8.48).
2. Generate a set of Thompson samples {𝑥𝑖∗ }.
3. For each sample 𝑥 ,∗ derive the joint distribution of the function value,
gradient, and Hessian at 𝑥 .∗ Let z∗ = (𝜙 ,∗ h∗, upper H∗, ∇∗ ) represent a vec-
tor comprising these random variables, with those that will be subjected
to expectation propagation first. We compute:

𝑝 (z∗ | 𝑥 ,∗ D) = N (z∗ ; 𝝁 ,∗ 𝚺∗ ).

47 In the interest of numerical stability, the in- Find the marginal belief over (𝜙 ,∗ h∗ ) and use Gaussian expectation prop-
verse of V∗ should not be stored directly; for agation to approximate the posterior after incorporating the factors (8.47,
relevant practical advice see:
8.49). Let vectors 𝝁˜ and 𝝈˜ 2 denote the site parameters at termination.
c. e. rasmussen and c. k. i. williams (2006). Finally, precompute47
"  # −1 "  #
Gaussian Processes for Machine Learning. mit
Press. 
𝚺˜ 0 −1 𝝁˜
j. p. cunningham et al. (2011). Gaussian Prob-
−1
V∗ = 𝚺 + ∗
; ∗
𝜶 = V∗ −𝝁 ,∗
0 0 0
abilities and Expectation Propagation. arXiv:
1111.6832 [stat.ML].
where 𝚺˜ = diag 𝝈˜ 2. These quantities do not depend on 𝑥 and will be
repeatedly reused during prediction.
After completing the preparations above, suppose a proposed ob-
servation location 𝑥 is given. For each sample 𝑥 ,∗ we compute the joint
distribution of 𝜙 and 𝜙 ∗ :

𝑝 (𝜙, 𝜙 ∗ | 𝑥, 𝑥 ,∗ D) = N (𝜙, 𝜙 ∗ ; 𝝁, 𝚺),

and derive the approximate posterior given the exact gradient (8.42)
and off-diagonal Hessian (8.45) observations and the factors (8.47, 8.49).
Defining  >
k

K = > = cov [𝜙, 𝜙 ∗ ] >, z∗ ,
k∗
48 This can be derived by marginalizing z∗ ac- the desired distribution is N (𝜙, 𝜙 ∗ ; m, S), where:48
cording to its approximate posterior from step
   2 
3 above: 𝑚 𝜍 𝜌
∫ m = ∗ = 𝝁 + K𝜶 ∗ S= = 𝚺 − KV∗−1 K>. (8.51)
𝑝 (𝜙, 𝜙 ∗ | z∗ ) 𝑝 (z∗ ) dz∗.
𝑚 𝜌 𝜍 ∗2

We now apply the prediction constraint (8.50) with one final step of
expectation propagation. Define (b.7, b.11–b.12):

𝜇¯ = 𝑚 − 𝑚 ∗ ; 𝜎¯ 2 = 𝜍 2 − 2𝜌 + 𝜍 ∗2 ;
  −1
𝜇¯ 𝜙 (𝑧) 𝜎¯ 𝜙 (𝑧)
𝑧=− ; 𝛼 =− ; 𝛾 =− +𝑧 .
𝜎¯ Φ(𝑧)𝜎¯ 𝛼 Φ(𝑧)

186
 8.9. mutual information with 𝑓 ∗

The final approximation to the predictive variance of 𝜙 given 𝑥 ∗ is

𝜎∗2 = 𝜍 2 − (𝜍 2 − 𝜌) 2 /𝛾, (8.52)

from which we can compute the contribution to the acquisition function

from this sample with (8.39–8.40).
Although it may seem unimaginable at this point, in Euclidean do- gradient of predictive entropy search
mains we may compute the gradient of the predictive entropy search acquisition function: § c.3, p. 311
acquisition function; see the accompanying appendix for details.

8.9 mutual information with 𝑓 ∗

Finally, we consider the computation of the mutual information between mutual information with 𝑓 ∗ : § 7.6, p. 140
the observed value 𝑦 and the value of the global maximum 𝑓 ∗ (8.53).
Several authors have considered this acquisition function in its predictive
form (8.35), which is the most convenient choice for Gaussian process
models:
 
𝛼 𝑓 ∗ (𝑥; D) = 𝐻 [𝑦 | 𝑥, D] − 𝔼 𝐻 [𝑦 | 𝑥, 𝑓,∗ D] | 𝑥, D . (8.53)

Unfortunately, this expression cannot be computed exactly due to the 49 z. wang and s. jegelka (2017). Max-value En-
complexity of the distribution 𝑝 (𝑓 ∗ | D); see figure 8.17 for an example. tropy Search for Efficient Bayesian Optimiza-
tion. icml 2017.
However, several effective approximations have been proposed, including
max-value entropy search (mes)49 and output-space predictive entropy 50 m. w. hoffman and z. ghahramani (2015).
Output-Space Predictive Entropy Search for
search (opes).50 Flexible Global Optimization. Bayesian Opti-
The issues we face in estimating (8.53), and the strategies we use to mization Workshop, neurips 2015.
overcome them, largely mirror those in predictive entropy search. To
begin, the first term is the differential entropy of a Gaussian and may be
computed exactly:

𝐻 [𝑦 | 𝑥, D] = 1
2 log(2𝜋𝑒𝑠 2 ). (8.54)

The second term, however, presents some challenges, and the available
approximations to (8.53) diverge in their estimation approach. We will
discuss the mes and opes approximations in parallel, as they share the
same basic strategy and only differ in some details along the way.

Approximating expectation with respect to 𝑓 ∗

The first complication in evaluating the second term of (8.53) is that we
must compute an expectation with respect to 𝑓 ∗. Although Thompson
sampling and simple Monte Carlo approximation is one way forward,
we can exploit the fact that 𝑓 ∗ is one dimensional to pursue more sophis-
ticated approximations. One convenient and rapidly converging strategy
is to design 𝑛 samples {𝑓𝑖∗ }𝑛𝑖=1 to be equally spaced quantiles of 𝑓 ,∗ then
use the familiar estimator

  1 ∑︁𝑛
𝔼 𝐻 [𝑦 | 𝑥, 𝑓,∗ D] | 𝑥, D ≈ 𝐻 [𝑦 | 𝑥, 𝑓𝑖∗, D]. (8.55)
𝑛 𝑖=1
Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 187
 computing policies with gaussian processes

This represents a quasi-Monte Carlo51 approximation that converges

51 r. e. caflisch (1998). Monte Carlo and quasi-
Monte Carlo methods. Acta Numerica 7:1–49. considerably faster than simple Monte Carlo integration.52 Further, this
52 The chosen samples are the first 𝑛 points from
scheme only requires the ability to estimate quantiles of 𝑓 ∗.
a base-𝑛 van de Corput sequence:

j. g. van de corput (1935). Verteilungsfunktio- Approximating quantiles of 𝑓 ∗

nen: Erste Mitteilung. Proceedings of the Konin-
klijke Nederlandse Akademie van Wetenschap- wang and jegelka proposed estimating the quantiles of 𝑓 ∗ via a conve-
pen 38:813–821 nient analytic approximation. To build this approximation, they proposed
mapped to the desired distribution via the in-
selecting a set of representer points 𝝃 and approximating the distribu-
verse cdf. These points have the minimum tion of 𝑓 ∗ with that of the maximum function value restricted to these
possible discrepancy for a set of size 𝑛. locations. Let 𝝓 = 𝑓 (𝝃 ) and define the random variable 𝜙 ∗ = max 𝝓. We
estimate:
𝑝 (𝑓 ∗ | D) ≈ 𝑝 (𝜙 ∗ | 𝝃 , D).
53 c. e. clark (1961). The Greatest of a Finite Set Unfortunately, the distribution of the maximum of dependent Gaus-
of Random Variables. Operations Research 9(2): sian random variables is intractable in general, even if the dimension is
145–162.
finite,53, 54 so we must resort to further approximation.
54 a. m. ross (2010). Computing Bounds on the We could proceed in several ways, but wang and jegelka propose us-
Expected Maximum of Correlated Normal
Variables. Methodology and Computing in Ap- ing an exhaustive, dense set of representer points (for example, covering
plied Probability 12(1):111–138. the domain with a low-discrepancy sequence) and then making the sim-
plifying assumption that the associated function values are independent.
If the marginal belief at each representer point is

𝑝 (𝜙𝑖 | 𝜉𝑖 , D) = N (𝜙𝑖 ; 𝜇𝑖 , 𝜎𝑖2 ),

then we may approximate the cumulative distribution function of 𝜙 ∗

with a product of normal cdfs:
Ö  𝑧 − 𝜇𝑖 
Pr(𝜙 ∗ < 𝑧 | 𝝃 , D) ≈ Φ . (8.56)
𝑖
𝜎𝑖

We can now estimate the quantiles of 𝑓 ∗ via numerical inversion of (8.56);

55 r. p. brent (1973). Algorithms for Minimization Brent’s55 method applied to the log cdf would offer rapid convergence.
Without Derivatives. Prentice–Hall. [chapter The resulting approximation for our example scenario is shown in figure
4]
8.17 using 100 equally spaced representer points covering the domain. In
56 d. slepian (1962). The One-Sided Barrier Prob- general, the independence assumption tends to overestimate the maximal
lem for Gaussian Noise. The Bell System Tech-
nical Journal 41(2):463–501. value, a finding wang and jegelka explained heuristically by appealing
to slepian’s lemma.56, 57
57 This result requires the posterior covariance
function to be positive everywhere, which is A diametrically opposing alternative to using dense representer
not guaranteed even if true for the prior co- points with a crude approximation (8.56) would be using a few, carefully
variance function. However, it is “usually true” chosen representer points with a better approximation to the distribu-
for typical models in high-dimensional spaces.
tion of 𝜙 ∗. For example, we could generate a set of Thompson samples,
𝜉𝑖 ∼ 𝑝 (𝑥 ∗ | D), then approximate the quantiles of 𝜙 ∗ by repeatedly sam-
pling their corresponding function values 𝝓. Figure 8.17 shows such an
approximation for our example scenario using 100 Thompson samples.
The agreement with the true distribution is excellent, but the computa-
tional cost was significant compared to the independent approximation.
However, in high-dimensional spaces where dense coverage of the do-
main is not possible, such a direct approach may become appealing.

188
 8.9. mutual information with 𝑓 ∗

𝑝 (𝑓 ∗ | D) independent approximation dependent approximation

𝑓∗ 𝑓∗ 𝑓∗
max y max y max y

Figure 8.17: Left: the true distribution 𝑝 (𝑓 ∗ | D) for our running scenario, estimated using exhaustive sampling. Middle:
wang and jegelka’s approximation to 𝑝 (𝑓 ∗ | D) (8.56) using a grid of 100 equally spaced representer points.
Right: an approximation to 𝑝 (𝑓 ∗ | D) from sampling the function values at the 100 Thompson samples from
figure 8.11.

Predictive entropy with exact observations

Regardless of the exact estimator we use to approximate the second term
of the acquisition function, we will need to compute the entropy of the
predictive distribution for 𝑦 given the optimal value 𝑓 ∗ :
∫
𝑝 (𝑦 | 𝑥, 𝑓, D) = 𝑝 (𝑦 | 𝑥, 𝜙) 𝑝 (𝜙 | 𝑥, 𝑓,∗ D) d𝜙 .
∗
(8.57)

Remarkably, unlike in the case of predictive entropy search, the latent

predictive distribution of 𝜙 given 𝑓 ∗ can be derived easily as a normal
distribution with upper tail truncated at 𝑓 ∗ :


𝑝 (𝜙 | 𝑥, 𝑓,∗ D) = TN 𝜙; 𝜇, 𝜎 ,2 (−∞, 𝑓 ∗ ) . (8.58)
Even more remarkably, the entropy of this distribution can be computed
in closed form:
 
1 2 2
𝜙 (𝑧) 𝑓∗ −𝜇
𝐻 [𝜙 | 𝑥, 𝑓, D] = log 2𝜋𝑒𝜎 Φ(𝑧) − 𝑧
∗
; 𝑧= . (8.59)
2 Φ(𝑧) 𝜎
In the absence of observation noise, this result is sufficient to realize
a complete algorithm by combining (8.59) with (8.53–8.55). After simpli-
fication, this yields the final mes approximation, which ignores the effect
of observation noise in the predictive distribution:
𝑛  
1 ∑︁ 𝜙 (𝑧𝑖 ) 2 𝑓∗ −𝜇
𝛼 mes (𝑥; D) = 𝑧𝑖 − log Φ(𝑧𝑖 ) ; 𝑧𝑖 = 𝑖 . (8.60)
2𝑛 𝑖=1 Φ(𝑧𝑖 ) 𝜎
Figure 8.18 illustrates this approximation for our running example using
the independent approximation to 𝑝 (𝑓 ∗ | D) (8.56). The approximation
is faithful and induces a near-optimal decision; any (slight) inaccuracy is
entirely due to the independence assumption.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 189

 computing policies with gaussian processes

mutual information with 𝑓,∗ 𝛼 𝑓 ∗ next observation location

mes approximation (8.60)

Figure 8.18: An approximation to the mutual information between the observed value 𝑦 and the value of the global
optimum 𝛼 𝑓 ∗ = 𝐼 (𝑦; 𝑓 ∗ | 𝑥, D) for our running example using the independent approximation to 𝑝 (𝑓 ∗ | D)
(8.56) and numerical integration.

Approximating predictive entropy with noisy observations

direct computation of predictive entropy In the case of additive Gaussian noise, however, the predictive distribu-
tion of 𝑦 (8.57) becomes the convolution of a centered normal distribution
and a truncated normal distribution; see figure 8.19 for an illustration
58 s. turban (2010). Convolution of a truncated of a particularly noisy scenario. turban provides a closed form for the
normal and a centered normal variable. Tech- resulting probability density function:58
nical report. Columbia University.
   
𝛼 (𝑦) − 𝑦 + 𝑓 ∗ (𝑦 − 𝜇) 2
∗
𝑝 (𝑦 | 𝑥, 𝑓, D) = 𝛾 Φ exp − ,
𝛽 2𝑠 2
where
𝜎𝑛2 (𝑦 − 𝜇) 𝜎𝑛 𝜎 1
𝛼 (𝑦) = ; 𝛽= ; 𝛾=√ ,
𝑠2 𝑠 2𝜋𝑠 Φ(𝑧)

and 𝑧 is defined in (8.59). Unfortunately this distribution does not admit

a closed-form expression for its entropy, although we can approximate
the entropy effectively via numerical quadrature.
analytic approximation to predictive entropy Alternatively, we can appeal to analytic approximation to estimate
the predictive entropy, and both wang and jegelka and hoffman and
59 However, the closed form for the predictive ghahramani take this approach.59 The mes approximation simply ig-
distribution above was not discussed by either nores the observation noise in the acquisition function (8.60) and instead
and perhaps deserves further consideration.
computes the mutual information between the latent function value 𝜙
and 𝑓 .∗ This approach overestimates the true mutual information by
assuming exact observations, but serves as a reasonable approximation
when the signal-to-noise ratio is high. This approximation is illustrated
in figure 8.20 for a scenario with relatively high noise; although the
approximation is not perfect, the chosen point is close to optimal.

190
 8.9. mutual information with 𝑓 ∗

𝑝 (𝜙 | 𝑥, 𝑓,∗ D) Gaussian ep approximation

𝑝 (𝑦 | 𝑥, 𝑓,∗ D)

𝑦 𝑦
𝑓∗ 𝑓∗

Figure 8.19: Left: an example of the latent predictive distribution 𝑝 (𝜙 | 𝑥, 𝑓,∗ D), which takes the form of a truncated
normal distribution (8.58), and the resulting predictive distribution 𝑝 (𝑦 | 𝑥, 𝑓,∗ D), which is a convolution with
a centered normal distribution accounting for Gaussian observation noise. Right: a Gaussian expectation
propagation approximation to the predictive distribution (8.62).

hoffman and ghahramani instead approximate the latent predictive truncating a variable with ep: § b.2, p. 305
distribution (8.58) using Gaussian expectation propagation:

𝑝 (𝜙 | 𝑥, 𝑓,∗ D) ≈ N (𝜙; 𝜇∗, 𝜎∗2 ),

where
 
𝜙 (𝑧) 𝜙 (𝑧) 2
𝜎∗2 2
= 𝜎 1−𝑧 − (8.61)
Φ(𝑧) Φ(𝑧) 2

is the variance of the truncated normal latent predictive distribution

(8.58) and 𝑧 is defined in (8.59). We may now approximate the differential
entropy of 𝑦 with

var[𝑦 | 𝑥, 𝑓,∗ D] ≈ 𝜎∗2 + 𝜎𝑛2 = 𝑠 ∗2 ;

(8.62)
𝐻 [𝑦 | 𝑥, 𝑓,∗ D] ≈ 1
2 log(2𝜋𝑒𝑠 ∗2 ).

This is similar to the strategy used in predictive entropy search, al-

though computing the approximate latent posterior is considerably easier
in this case, only requiring the single factor [𝜙 < 𝑓 ∗ ]. Figure 8.19 shows
the resulting Gaussian approximation to the predictive distribution for
an example point; the approximation is excellent. Estimating the expec-
tation over 𝑓 ∗ with (8.55) and simplifying, the final opes approximation
to (8.53) is
1 ∑︁
𝑛
𝛼 opes (𝑥; D) = log 𝑠 − log 𝑠 ∗𝑖 , (8.63)
𝑛 𝑖=1
where 𝑠 ∗𝑖 is the approximate predictive standard deviation corresponding
to the sample 𝑓𝑖∗. This takes the same form as the predictive entropy
search approximation to the mutual information with 𝑥 ∗ (8.41), although
the approximations to the predictive distribution are of course different.
The opes approximation (8.63) is shown for a high-noise scenario in
figure 8.20; the approximation is almost perfect and in this case yields
the optimal decision.
Both the mes and opes approximations to the mutual information
can be differentiated with respect to the proposed observation location.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 191

computing policies with gaussian processes

mutual information with 𝑓,∗ 𝛼 𝑓 ∗ next observation location

mes approximation (8.60)

opes approximation (8.62)

Figure 8.20: The mes and opes approximations to the mutual information with 𝑓 ∗ for an example high-noise scenario
with unit signal-to-noise ratio. Both use the independent representer point approximation to 𝑝 (𝑓 ∗ | D) (8.56).

For the former, we simply differentiate (8.60):

  
1 ∑︁ 𝜙 (𝑧𝑖 ) 𝜕𝜇
𝑛
𝜕𝛼 mes 𝜕𝜎 𝜙 (𝑧𝑖 )
= + 𝑧𝑖 1 + 𝑧𝑖 + 𝑧𝑖2 ;
𝜕𝑥 2𝑛𝜎 𝑖=1 Φ(𝑧𝑖 ) 𝜕𝑥 𝜕𝑥 Φ(𝑧𝑖 )

note the {𝑓𝑖∗ } samples will in general not depend on 𝑥, so we do not need
to worry about any dependence of their distribution on the observation
gradient of opes acquisition function: § c.3, location. The details for the opes approximation are provided in the
p. 311 appendix.

8.10 averaging over a space of gaussian processes

Throughout this chapter we have assumed our belief regarding the objec-
tive function is given by a single Gaussian process. However, especially
when our dataset is relatively small, we may seek robustness to model
model averaging: § 4.4, p. 74 misspecification by averaging over multiple plausible models. Here we
will provide some guidance for policy computation when performing
Bayesian model averaging over a space of Gaussian processes. Although
this may seem straightforward, there is some nuance involved.
Below we will consider a space of Gaussian processes indexed by a
vector of hyperparameters 𝜽 , which determines both the moments the
objective function prior and any relevant parameters of the observation

192
 8.10. averaging over a space of gaussian processes

noise process. Bayesian model averaging over this space yields marginal
posterior and predictive distributions:
∫
𝑝 (𝑓 | D) = 𝑝 (𝑓 | D, 𝜽 ) 𝑝 (𝜽 | D) d𝜽 ; (8.64)
∫
𝑝 (𝑦 | 𝑥, D) = 𝑝 (𝑦 | 𝑥, D, 𝜽 ) 𝑝 (𝜽 | D) d𝜽, (8.65)

which are integrated against the model posterior 𝑝 (𝜽 | D) (4.7). Both of model posterior, 𝑝 (𝜽 | D)
these distributions are in general intractable, but we developed several
viable approximations in chapter 3, all of which approximate the objective
function posterior (8.64) with a mixture of Gaussian processes and the
posterior predictive distribution (8.65) with a mixture of Gaussians.

Noiseless expected improvement and probability of improvement

When observations are exact, the marginal gain in utility underlying
both expected improvement and probability of improvement depends
only on the objective function value 𝜙 and the value of the incumbent
𝜙 ∗:
Δei (𝑥, 𝜙) = max(𝜙 − 𝜙 ,∗ 0); Δpi (𝑥, 𝜙) = [𝜙 > 𝜙 ∗ ].
As a result, the formulas we derived for these acquisition functions given
a gp belief on the objective (8.9, 8.21) depend only on the moments of the 60 In fact, a Gaussian process belief is not required
(Gaussian) predictive distribution 𝑝 (𝜙 | 𝑥, D).60 With a Gaussian mixture at all, only Gaussian predictive distributions.
This will be important in the next section.
approximation to the predictive distribution, the expected marginal gain,
∫
𝛼 (𝑥; D) = Δ(𝑥, 𝜙) 𝑝 (𝜙 | 𝑥, D) d𝜙,

is simply a weighted combination of these results by linearity of expec-

tation.

Model-dependent utility functions

For the remaining decision-theoretic acquisition functions, noisy ex-
pected improvement and probability of improvement, knowledge gra-
dient, and mutual information with any relevant random variable 𝜔,
the situation is somewhat more complicated, because the underlying
utility functions depend on the model of the objective function. The first dependence on objective function model
three depend on the posterior mean function, and the last depends on
the posterior belief 𝑝 (𝜔 | D), both of which are induced by our belief
regarding the objective function. To make this dependence explicit, for
a model 𝜽 in our space of interest, let us respectively notate the utility,
marginal gain in utility, and expected marginal gain in utility with:

𝑢 (D; 𝜽 ); Δ(𝑥, 𝑦; 𝜽 ); 𝛼 (𝑥; D, 𝜽 ).

There are two natural ways we might address this dependence when
averaging over models. One is to seek to maximize the expected utility

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 193

 computing policies with gaussian processes

average model-conditional utility, 𝔼𝑢 of the data, averaged over the choice of model:
∫
𝔼𝑢 (D) = 𝑢 (D; 𝜽 ) 𝑝 (𝜽 | D) d𝜽 . (8.66)

marginal gain in 𝔼𝑢, 𝔼Δ Writing the marginal gain in expected utility as 𝔼Δ, we may derive an
expected marginal gain in 𝔼𝑢, 𝔼𝛼 acquisition function via one-step lookahead:
∫
𝔼𝛼 (𝑥; D) = 𝔼Δ(𝑥, 𝑦) 𝑝 (𝑦 | 𝑥, D) d𝑦
∫ ∫ 
= Δ(𝑥, 𝑦; 𝜽 ) 𝑝 (𝑦 | 𝑥, D, 𝜽 ) d𝑦 𝑝 (𝜽 | D) d𝜽
∫
= 𝛼 (𝑥; D, 𝜽 ) 𝑝 (𝜽 | D) d𝜽 . (8.67)

As hinted by its notation, this is simply the expectation of the conditional

acquisition functions, which we can approximate via standard methods.
Although this approach is certainly convenient, it may overestimate
optimization progress as utility is only measured under the assumption
of a perfectly identified model – the utility function is “blind” to model
uncertainty. An arguably more appealing alternative is to evaluate utility
with respect to the marginal objective function model (8.64) from the
start, defining simple and global reward with respect to the marginal
marginal posterior mean posterior mean: ∫
𝜇D (𝑥; 𝜽 ) 𝑝 (𝜽 | D) d𝜽, (8.68)

marginal belief about 𝜔 and information gain about 𝜔 with respect to its marginal belief:
∫
𝑝 (𝜔 | D, 𝜽 ) 𝑝 (𝜽 | D) d𝜽 . (8.69)

utility of marginal model, 𝑢𝔼 Let us notate a utility function defined in this manner with 𝑢𝔼(D), con-
(expected) marginal gain in utility of trasting with its post hoc averaging equivalent, 𝔼𝑢 (D) (8.66). Similarly,
marginal model: Δ𝔼, 𝛼𝔼 let us notate its marginal gain with Δ𝔼 and its expected marginal gain
with: ∫
𝛼𝔼(𝑥; D) = Δ𝔼(𝑥, 𝑦) 𝑝 (𝑦 | 𝑥, D) d𝑦. (8.70)

Example and discussion

We may shed some light on the differences between these approaches
with a barebones example. Let us work with the knowledge gradient,
61 These models may be interpreted as degener- and consider a pair of simple models for an objective function on the
ate Gaussian processes with covariance 𝐾 ≡ 0. interval X = [−1, 1]: either 𝑓 = 𝑥 or 𝑓 = −𝑥, with equal probability.61
The model-conditional global reward in either case is 1, and thus
the expected utility (8.66) does not depend on the data: 𝔼𝑢 (D) ≡ 1. What
a strange set of affairs – although we know the maximal value of the
objective a priori, we need data to tell us where it is! For this particular
model space, an optimal recommendation for either model is maximally
suboptimal if that model is incorrect.

194
 8.10. averaging over a space of gaussian processes

In contrast, the global reward of the marginal model does depend on

the data via the model posterior. The global reward is monotonic in the
probability of the model most favored by the data, 𝜋:
𝑢𝔼(D) = 2𝜋 − 1.
A priori, the utility is 𝑢𝔼(∅) = 0 – the marginal mean function is 𝜇 ≡ 0,
and no compelling recommendation is possible until we determine the
correct model.
As the expected utility 𝔼𝑢 (8.66) is independent of the data, it does
not lead to an especially insightful policy. In contrast, the marginal gain
in 𝑢𝔼 (8.70) can differentiate potential observation locations via their
expected impact on the model posterior. We illustrate this acquisition
function in the margin given an empty dataset and assuming moderate Above: the two possible objectives for our
additive Gaussian noise. As one might hope, it prefers evaluating on the comparison of the 𝔼𝑢 and 𝑢𝔼. Below: the
expected marginal gain in 𝑢𝔼, which prefers
boundary of the domain, where observations are expected to provide sampling on the boundary to reveal more
more information regarding the model and thus greater improvement in information regarding the model. The
model-marginal utility. expected marginal gain in 𝔼𝑢 is constant.

Computing expected gain in marginal utility

Unfortunately, the expected gain in utility for the marginal model (8.70)
does not simplify as before (8.66), as we must account for the effect of the
observed data on the model posterior in the utility function. However,
we may sketch a Monte Carlo approximation using samples from the
model posterior, {𝜽𝑖 } ∼ 𝑝 (𝜽 | D).
For utility functions based on the marginal posterior mean (8.68),
the resulting approximation to the expected marginal gain (8.70) is a
weighted sum of Gaussian expectations of Δ𝔼, each of which we may
approximate via Gauss–Hermite quadrature. To approximate Δ𝔼 for a Gauss–Hermite quadrature: § 8.5, p. 171
putative observation (𝑥, 𝑦), a simple sequential Monte Carlo approxima-
tion to the updated model posterior would reweight each sample 𝜽𝑖 by
𝑝 (𝑦 | 𝑥, D, 𝜽𝑖 ), then approximate the updated marginal posterior mean
by a weighted sum.
To approximate the predictive form of mutual information with 𝑥 ∗
or 𝑓 ∗ (8.36, 8.53):
 
𝐻 [𝑦 | 𝑥, D] − 𝔼𝜔 𝐻 [𝑦 | 𝑥, 𝜔, D] | 𝑥, D ,
we first note that the first term is the entropy of a Gaussian mixture, 62 This requires one final layer of approximation,
which we can approximate via quadrature. We may approximate the which is produced as a byproduct of expecta-
tion propagation in the case of 𝑥 ∗ (b.6), and
expectation in the second term via Thompson sampling (see below); may be dealt with without much trouble in
for each of these samples, we may approximate the updated predictive the case of the univariate random variable 𝑓 ∗.
distribution as before, reweighting the mixture by 𝑝 (𝜔 | D, 𝜽 ).62

Upper confidence bound and Thompson sampling

We may compute any desired upper confidence bound of a Gaussian
process mixture at 𝑥 by bisecting the cumulative distribution function
of 𝜙, which is a weighted sum of Gaussian cdfs.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 195

 computing policies with gaussian processes

Finally, to perform Thompson sampling from the marginal posterior,

we first sample from the model posterior, 𝜽 ∼ 𝑝 (𝜽 | D); the conditional
Thompson sampling for gps: § 8.7, p. 176 posterior 𝑝 (𝑓 | D, 𝜽 ) is then a gp, and we may proceed via the previous
discussion.

8.11 alternative models: bayesian neural networks, etc.

Although Gaussian processes are without question the most prominent
objective function model used in Bayesian optimization, we are of course
free to use any other model when prudent. Below we briefly outline some
notable alternative model classes that have received some attention in
the context of Bayesian optimization and comment on any issues arising
in computing common policies with these surrogates.

Random forests
63 l. breiman (2001). Random Forests. Machine Random forests63 are a popular model class renown for their excellent
Learning 45(1):5–32. off-the-shelf performance,64 offering good generalization, strong resis-
64 m. fernández-delgado et al. (2014). Do we tance to overfitting, and efficient training and prediction. Of particular
Need Hundreds of Classifiers to Solve Real relevance for optimization, random forests are adept at handling high-
World Classification Problems? Journal of Ma-
chine Learning Research 15(90):3133–3181. dimensional data and categorical and conditional features, and may be
a better choice than Gaussian processes for objectives featuring any of
these characteristics.
Algorithm configuration is one setting where these capabilities are
critical: complex algorithms such as compilers or sat solvers often have
complex configuration schemata with many mutually dependent param-
eters, and it can be difficult to build nontrivial covariance functions for
such inputs. Random forests require no special treatment in this setting
65 f. hutter et al. (2014). Algorithm runtime pre- and have delivered impressive performance in predicting algorithmic
diction: Methods & evaluation. Artificial Intel- performance measures such as runtime.65 They are thus a natural choice
ligence 206:79–111.
for Bayesian optimization of these same measures.66
66 f. hutter et al. (2011). Sequential Model-Based Classical random forests are not particularly adept at quantifying un-
Optimization for General Algorithm Configu-
ration. lion 5. certainty in predictions off-the-shelf. Seeking more nuanced uncertainty
quantification, hutter et al. proposed a modification of the vanilla model
wherein leaves store both the mean (as usual) and the standard deviation
of the training data terminating there.65 We then estimate the predictive
distribution with a mixture of Gaussians with moments corresponding
67 hutter et al. then fit a single Gaussian distri- to the predictions of the member trees.67, 68 Figure 8.21 compares the
bution to this mixture via moment matching, predictions of a Gaussian process and a random forest model on a toy
although this is not strictly necessary.
dataset. Although they differ in their extrapolatory behavior, the models
68 A similar approach can also be used to esti- make very similar predictions otherwise.
mate arbitrary predictive quantiles:
To realize an optimization policy with a random forest, hutter et
n. meinshausen (2006). Quantile Regression al. suggested approximating acquisition functions depending only on
Forests. Journal of Machine Learning Research marginal predictions – such as (noiseless) expected improvement or
7(35):983–999.
probability of improvement – by simply plugging this Gaussian approxi-
mation into the expressions derived in this chapter (8.9, 8.22). Either can
be computed easily from a Gaussian mixture predictive distribution as
well due to linearity of expectation.

196
 8.11. alternative models: bayesian neural networks, etc.

Figure 8.21: The predictions of a Gaus-

sian process model (above)
and a random forest model
comprising 100 regression
trees (below) for an exam-
ple dataset; the credible inter-
vals of the latter are not sym-
metric as the predictive dis-
tribution is estimated with a
Gaussian mixture.

Density ratio estimation

bergstra et al. described a lightweight Bayesian optimization algorithm 69 j. bergstra et al. (2011). Algorithms for Hyper-
that operates by maximizing probability of improvement (7.3) via a re- Parameter Optimization. neurips 2011.
duction to density ratio estimation.69, 70 70 Although the source material claims that ex-
We begin each iteration of this algorithm by choosing some reference pected improvement is maximized, this was
the result of a minor mathematical error.
value 𝑦 ∗ that we wish to exceed with our next observation; we will
71 Although we have spent considerable effort
then select the next observation location to maximize the probability of arguing against maximizing “improvement”
improvement over this threshold:71 over a noisy observation – at least in the pres-
ence of high noise – we will see that this choice
𝛼 pi (𝑥; D, 𝑦 ∗ ) = Pr(𝑦 > 𝑦 ∗ | 𝑥, D). (8.71) provides considerable computational benefit.

We will discuss the selection of the improvement target 𝑦 ∗ shortly.

We now consider two conditional probability density functions de-
pending on this threshold:

𝑔(𝑥) = 𝑝 (𝑥 | 𝑦 > 𝑦,∗ D);

(8.72)
ℓ (𝑥) = 𝑝 (𝑥 | 𝑦 ≤ 𝑦,∗ D);

that is, 𝑔 is the probability density of observation locations exceeding

the threshold, and ℓ of locations failing to. Note that both densities are
proportional to quantities related to the probability of improvement:

𝑔(𝑥) ∝ Pr(𝑦 > 𝑦 ∗ | 𝑥, D) 𝑝 (𝑥) = 𝛼 pi 𝑝 (𝑥);

(8.73)
ℓ (𝑥) ∝ Pr(𝑦 ≤ 𝑦 | 𝑥, D) 𝑝 (𝑥) = (1 − 𝛼 pi ) 𝑝 (𝑥),
∗

where 𝑝 (𝑥) is an arbitrary prior density that we will dispose of shortly.

bergstra et al. now define an acquisition function as the ratio of
these densities:
𝑔(𝑥) 𝛼 pi
𝛼 (𝑥; D) = ∝ , (8.74)
ℓ (𝑥) 1 − 𝛼 pi
where the prior density 𝑝 (𝑥) appearing in (8.73) cancels.72 As this ratio 72 There is, however, a tacit assumption that the
is monotonically increasing with the probability of improvement (8.71), prior density has support over the whole do-
main to enable this cancellation.
maximizing the density ratio yields the same policy.
To proceed, we require some mechanism for estimating this density
ratio (8.74). bergstra et al. appealed to density estimation, constructing

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 197

 computing policies with gaussian processes

Figure 8.22: bergstra et al.’s Parzen es-

timation optimization pol- 𝑦∗
icy. The top panel shows
an example dataset along
with a threshold 𝑦 ∗ set to
be the 85th percentile of the
observed values. The next
two panels illustrate the cen-
tral kernel density estimates 𝑔(𝑥)
(8.72) for the density of ob-
servation locations above
and below this threshold.
The “wiggliness” of these es-
timates stems from the ker- ℓ (𝑥)
nel bandwidth scheme pro-
posed by bergstra et al. The
bottom panel shows the ra-
tio of these densities, which
is monotonic in the probabil- 𝑔(𝑥)/ℓ (𝑥)
ity of improvement over 𝑦 ∗
(8.74).

kernel (Parzen) density estimates73 of 𝑔 and ℓ and directly maximizing

73 b. w. silverman (1986). Density Estimation for
Statistics and Data Analysis. Chapman & Hall. their ratio (8.74) to realize a policy.74 One advantage of this scheme is
that the cost of computing the policy only scales linearly with respect to
74 The details of bergstra et al.’s scheme are
somewhat complicated due to the particular the number of observations. Figure 8.22 illustrates the key components
optimization problems they were considering. of this algorithm for a toy problem in one dimension.
Their focus was on hyperparameter tuning, An alternative to density estimation is to approximate the density
where conditional dependence among vari-
ables can lead to complex structure in the do-
ratio in (8.74) directly, an option promoted by tiao et al.75 Here we effec-
main. To address this dependence, bergstra tively reduce the problem to binary classification, building a probabilistic
et al. considered domains with tree structure, classifier to predict the binary label [𝑦 > 𝑦 ∗ ], then interpreting the pre-
where each node represents a variable that can dictions of this classifier in terms of the probability of improvement. Any
be assigned provided the assignments of its
parents. They then built separate conditional
probabilistic classifier could be used to this end, so the scheme offers a
density estimates for each variable appearing great deal of latitude in modeling beyond Gaussian processes or related
in the tree. models; in fact we do not need to model the objective function at all!
75 l. c. tiao et al. (2021). bore: Bayesian Opti- For either of these approaches to be feasible, we must have enough
mization by Density-Ratio Estimation. icml observations to either accurately estimate the required densities (8.72)
or to accurately train a classifier predicting [𝑦 > 𝑦 ∗ ]. bergstra et al.
2021.

addressed this issue by taking 𝑦 ∗ to be a relatively (but not exceedingly)

high quantile of the observed data, using the 85th percentile in their
experiments. tiao et al. provided some discussion on the role of 𝑦 ∗ in
role of improvement threshold in probability balancing exploration and exploitation, but ultimately used the 66th
of improvement: § 7.5, p. 133 percentile of the observed data in their experiments for simplicity.

Bayesian neural networks

Neural networks represent the state-of-the-art in numerous learning
tasks due to their flexibility in modeling and impressive predictive perfor-

198
 8.11. alternative models: bayesian neural networks, etc.

mance. For this reason, it tempting to use (Bayesian) neural networks as

objective function models in Bayesian optimization. However, there are
two obstacles that must be addressed in order to build a successful system.
First, the typically considerable cost of obtaining new data in Bayesian
optimization limits the amount of data that might be used to train a
neural network, imposing a ceiling on model complexity and perhaps
ruling out deep architectures. Second, in order to guide an optimization
policy, a neural network must yield useful estimates of uncertainty.
snoek et al. took an early step in this direction and reported success
with a relatively simple construction.76 The authors first trained a typical 76 j. snoek et al. (2015). Scalable Bayesian Opti-
(non-Bayesian) neural network using empirical loss minimization, then mization Using Deep Neural Networks. icml
2015.
replaced the final layer with Bayesian linear regression. The weights in
all but the final layer were fixed for the remainder of the procedure; the
resulting model can thus be interpreted as Bayesian linear regression
using neural basis functions. As this is in fact a Gaussian process,77 we 77 See the result in (3.6), setting 𝐾 ≡ 0 as there
may appeal to the computational details outlined in the remainder of is no nonlinear component in the model. One
beneficial side effect of this model is that the
this chapter to compute any policy we desire. cost of inference for Bayesian linear regression
A somewhat more involved (and, arguably, “more Bayesian”) ap- is only linear with the number of observations.
proach was proposed by springenberg et al.,78 who combined a para- 78 j. t. springenberg et al. (2016). Bayesian Op-
metric objective function model 𝑓 (𝑥; w) – the output of a neural network timization with Robust Bayesian Neural Net-
with input 𝑥 and weights w – with an additive Gaussian noise observa- works. neurips 2016.
tion model:

𝑝 (𝑦 | 𝑥, w, 𝜎) = N 𝑦; 𝑓 (𝑥; w), 𝜎 2 .
Bayesian inference proceeds by selecting a prior 𝑝 (w, 𝜎) and computing
the posterior from the observed data. The posterior is intractable, but 79 Following hutter et al.’s approach to opti-
springenberg et al. described a Hamiltonian Monte Carlo scheme for mization with random forests, springenberg
et al. fit a single Gaussian distribution to this
drawing samples {w𝑖 , 𝜎𝑖 }𝑠𝑖=1 from the posterior, from which we may form mixture via moment matching, but this is op-
a Gaussian mixture approximation to the predictive distribution:79 tional.

1 ∑︁

𝑠
𝑝 (𝑦 | 𝑥, D) ≈ N 𝑦; 𝑓 (𝑥; w𝑖 ), 𝜎𝑖2 .
𝑠 𝑖=1
80 This construction represents a realization of
This is sufficient to compute policies such as (noiseless) expected im- a manifold Gaussian process/deep kernel; see
provement and probability of improvement (8.9, 8.22). § 3.4, p. 59.
The impressive performance of modern deep neural networks is 81 r. gómez-bombarelli et al. (2018). Automatic
largely due to their ability to learn sophisticated feature representations Chemical Design Using a Data-Driven Contin-
uous Representation of Molecules. acs Central
of complex data, a process that requires enormous amounts of data and
Science 4(2):268–276.
may be out of reach in most Bayesian optimization settings. However,
82 Numerous example systems are outlined in:
when the domain consists of structured objects with sufficiently many
unlabeled examples available, one path forward is to train a generative b. l. hie and k. k. yang (2021). Adaptive ma-
latent variable model – such as a variational autoencoder or generalized chine learning for protein engineering. arXiv:
2106.05466 [q-bio.QM] [table 1].
adversarial network – using unsuperivsed (or semi-supervised) methods.
We can then perform optimization in the resulting latent space, for 83 j. j. irwin et al. (2020). zinc20 – A Free
Ultralarge-Scale Chemical Database for Lig-
example by simply constructing a Gaussian process over the learned and Discovery. Journal of Chemical Informa-
neural representation.80 This approach has notably proven useful in the tion and Modeling 60(12):6065–6073.
design of molecules81 and biological sequences.82 In both of these settings, 84 the uniprot consortium (2021). UniProt: the
enormous databases (on the order of hundreds of millions of examples) universal protein knowledgebase in 2021. Nu-
are available for learning a representation prior to optimization.83, 84 cleic Acids Research 49(d1):d480–d489.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 199

 computing policies with gaussian processes

Finally, neural networks have also received some attention in the

bandit literature as models for contextual bandits, which has lead to
85 d. zhou et al. (2020). Neural Contextual Ban- theoretical regret bounds.85, 86 riqelme et al. conducted an empirical
dits with ucb-based Exploration. icml 2020. “showdown” of various neural models in this setting.87 Although there
86 w. zhang et al. (2021). Neural Thompson Sam- was no clear “winner,” the authors concluded that “decoupled” models
pling. iclr 2021. – where uncertainty quantification is handled post hoc by feeding the
87 c. riqelme et al. (2018). Deep Bayesian Ban- output layer of a neural network into a simple model such as Bayesian
dits Showdown. iclr 2018. linear regression – tended to perform better than end-to-end training.

8.12 summary of major ideas

In this chapter we considered the computation of the popular optimiza-
tion policies described in the last chapter for Gaussian process models of
an objective function with an exact or additive Gaussian noise observa-
tion model. The acquisition functions for most of these policies represent
the one-step expected marginal gain to some underlying utility function:
∫
𝛼 (𝑥; D) = Δ(𝑥, 𝑦) N (𝑦; 𝜇, 𝑠 2 ) d𝑦,

where Δ(𝑥, 𝑦) is the gain in utility resulting from the observation (𝑥, 𝑦)
(8.27). When Δ is a piecewise linear function of 𝑦, this integral can be
resolved analytically in terms of the standard normal cdf. This is the
computation of expected improvement and case for the expected improvement and probability of improvement
probability of improvement: §§ 8.2–8.3, p. 167 acquisition functions, both with and without observation noise.
However, when Δ is a more complicated function of the putative
observation, we must in general rely on approximate computation to re-
solve this integral. When the predictive distribution is normal – as in the
approximate computation for one-step model class considered in this chapter – Gauss–Hermite quadrature pro-
lookahead: § 8.5, p. 171 vides a useful and sample-efficient approximation via a weighted average
computation of knowledge gradient: § 8.6, of carefully chosen integration nodes. This allows us to address some
p. 172 more complex acquisition functions such as the knowledge gradient.
The computation of mutual information with 𝑥 ∗ or 𝑓 ∗ entails an
expectation with respect to these random variables, which cannot be
approximated using simple quadrature schemes. Instead, we must rely on
Thompson sampling: § 8.7, p. 176 schemes such as Thompson sampling – a notable policy in its own right
approximate computation of mutual – to generate samples and proceed via (simple or quasi-) Monte Carlo
information: §§ 8.8–8.9, p. 180 integration, and in some cases, further approximations to the conditional
predictive distributions resulting from these samples.
Finally, Bayesian optimization is of course not limited to a single
Gaussian process belief on the objective function, which may be objec-
tionable even when Gaussian processes are the preferred model class
averaging over a space of gps: § 8.10, p. 192 due to uncertainty in hyperparameters or model structure. Averaging
over a space of Gaussian processes is possible – with some care – by
adopting a Gaussian process mixture approximation to the marginal
objective function posterior and relying on results from the single gp
alternatives to gps: § 8.11, p. 196 case. If desired, we may also abandon the model class entirely and com-
pute policies with respect to an alternative such as random forests or
Bayesian neural networks.

200
 9
IMPLEMEN TATION

There is a rich and mature software ecosystem available for Gaussian pro-
cess modeling and Bayesian optimization, and it is relatively easy to build
sophisticated optimization routines using off-the-shelf libraries. How-
ever, successful implementation of the underlying algorithms requires
attending to some nitty-gritty details to ensure optimal performance, and
what may appear to be simple equations on the page can be challenging
to realize in a limited-precision environment. In this chapter we will
provide a brief overview of the computational details that practitioners
should be aware of when designing Bayesian optimization algorithms,
even when availing themselves of existing software libraries.

9.1 gaussian process inference, scaling, and approximation

As is typical with nonparametric models, the computational cost of
Gaussian process inference grows (considerably!) with the number of
observations, and it is important to understand the nature of this growth
and be aware of methods for scaling to large-scale data when necessary.
The primary computational bottleneck in Gaussian process inference solving linear systems with respect to the
is solving systems of linear equations that scale with the number of observation covariance C
observed values. Consider the general case of exact inference where we
condition a Gaussian process GP (𝑓 ; 𝜇, 𝐾) on the observation of a length-
𝑛 vector of values y with marginal distribution and cross-covariance number of observed values, 𝑛 = |y|
function (2.6):
𝑝 (y) = N (y; m, C); 𝜅 (𝑥) = cov[y, 𝜙 | 𝑥]. (9.1)
The posterior is a Gaussian process with moments (2.10):
𝜇D (𝑥) = 𝜇 (𝑥) + 𝜅 (𝑥)> C−1 (y − m);
(9.2)
𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − 𝜅 (𝑥)> C−1 𝜅 (𝑥 0).
Evaluating either the posterior mean or the posterior covariance requires
solving a linear system with respect to the observation covariance C.
Inference with non-Gaussian observations using Monte Carlo sampling
or Gaussian approximate inference also entails solving linear systems
with respect to this matrix (2.35, 2.39–2.41).

Direct computation via Cholesky decomposition

neal outlined a straightforward implementation based on direct numer- 1 r. m. neal (1998). Regression and Classification
ical methods that suffices for the small-to-moderate datasets typical in Using Gaussian Process Priors. In: Bayesian
Statistics 6.
Bayesian optimization.1 We take advantage of the fact that C is sym-
metric and positive definite and precompute and store its Cholesky 2 g. h. golub and c. f. van loan (2013). Matrix
Computations. Johns Hopkins University Press.
factorization:2
[§ 4.2]
L = chol C; LL> = C.
This computation requires one-time O(𝑛 3 ) work and represents the bulk
of effort spent in this implementation. Note that despite having the same

This material will be published by Cambridge University Press as Bayesian Optimization. 201
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 implementation

asymptotic running time as the general-purpose lu decomposition, the

Cholesky factorization exploits symmetry to run twice as fast.
With the Cholesky factor in hand, we may solve an arbitrary linear
system Cx = b in time O(𝑛 2 ) via forward–backward substitution by
rewriting as LL> x = b and twice applying a triangular solver. We exploit
this fact to additionally precompute the vector

𝜶 = C−1 (y − m)

appearing in the posterior mean. After this initial precomputation of L

and 𝜶 , we may compute the posterior mean

𝜇D (𝑥) = 𝜇 (𝑥) + 𝜅 (𝑥) >𝜶

on demand in linear time and the posterior covariance

  >  −1 
𝐾D (𝑥, 𝑥 0) = 𝐾 (𝑥, 𝑥 0) − L−1𝜅 (𝑥) L 𝜅 (𝑥 0)
Í
3 The 2 𝑖 log 𝐿𝑖𝑖 term is the log determinant in quadratic time. We may also efficiently compute the log marginal
of C, where we have exploited the Cholesky likelihood (4.8) of the data in linear time:3
factorization and the fact that L is triangular.
 Í 
log 𝑝 (y | x) = − 21 (y − m) >𝜶 + 2 𝑖 log 𝐿𝑖𝑖 + 𝑛 log 2𝜋 .

Low-rank updates to the Cholesky factorization for sequential inference

Optimization is an inherently sequential procedure, where our dataset
grows incrementally as we gather new observations. In this setting, we
4 That is, a Gaussian process whose hyperpa- can accelerate sequential inference with a fixed Gaussian process4 by
rameters are fixed, rather than dependent on replacing direct computation of the Cholesky decomposition in favor of
the data.
a fast incremental updates to previously computed Cholesky factors.
For the sake of argument, suppose we have computed the Cholesky
factor L = chol C for a set of 𝑛 observations y with covariance C. Suppose
we then receive 𝑘 additional observations 𝝊, resulting in the augmented
observation vector  
0 y
y = .
𝝊
The covariance matrix of y0 is formed by appending the previous covari-
ance C with new rows/columns:
 
C X> 1
cov[y ] = C =
0 0
.
X1 X2

The Cholesky factor of the updated covariance matrix has the form
 
L 0
chol C =
0
.
𝚲1 𝚲2

Note that the upper-left block is simply the previously computed Cholesky
factor, which we can reuse. The new blocks may be computed as
 
𝚲1 = X1 L−> ; 𝚲2 = chol X2 − 𝚲1 𝚲> 1 .

202
 9.1. gaussian process inference, scaling, and approximation

The first of these blocks can be computed efficiently using a triangular

solver with the previous Cholesky factor, and the second block only
requires factoring a (𝑘 × 𝑘) matrix.
This low-rank update requires O(𝑘𝑛 2 ) work to compute, compared to
the O(𝑛 3 + 𝑘𝑛 2 ) work required to compute the Cholesky decomposition
of C0 from scratch; asymptotically, the low-rank update is O(𝑛/𝑘) times
faster. In particular, if we begin with an empty dataset and sequentially amortized analysis
apply this update for a total of 𝑛 observations (in any order and with
any number of observations at a time) the total cost of inference would
be O(𝑛 3 ). This is equivalent to the cost of one-time inference with the
full dataset, so the update scheme is as efficient as one could hope for.

Ill-conditioned covariance matrices

When an optimization policy elects to make an observation in the pursuit
of “exploitation,” the value observed is (by design!) highly correlated
with at least one existing observation. Although these decisions may be
well-grounded, the resulting highly correlated observations can wreak
havoc on numerical linear algebra routines.
To sketch the problems that may arise, consider the following sce-
nario. Suppose the covariance function is stationary, that observations
are corrupted by additive Gaussian noise with scale 𝜎𝑛2 , and that the
signal-to-noise ratio is large. Now, if two locations in the dataset corre-
spond to highly correlated values, the corresponding rows/columns in
the observation covariance matrix C will be nearly equal, and C will thus
be nearly singular – one or more eigenvalue will be near zero, and in
extreme cases, some may even become (numerically) negative. This poor
conditioning5 can cause loss of precision when solving a linear system 5 Numerical conditioning of C is usually evalu-
via the Cholesky decomposition, and a negative eigenvalue will cause ated by its condition number, defined in terms
of its singular values 𝝈 = {𝜎𝑖 }:
the Cholesky routine to fail altogether.
max 𝝈
When necessary, we may sidestep these issues with a number of 𝜅= .
min 𝝈
“tricks.” One simple solution is to add a small (in terms of 𝜎𝑛2 ) multiple
Given infinite precision, the singular values
of the identity to the observation covariance, replacing C by C + 𝜀 I. are equal to the (nonnegative) eigenvalues of
Numerically, this shifts the singular values of C by 𝜀, improving condi- C, but very poor conditioning can cause nega-
tioning; practically, this caps the signal-to-noise ratio by increasing the tive (numerical) eigenvalues.
noise floor. When high correlation is a result of small spatial separation, 6 That is, we replace observations of 𝑓 (𝑥) and
osborne et al. suggested replacing problematic observations of the objec- 𝑓 (𝑥 0 ) with one of 𝑓 (𝑥) and one of the direc-
tional derivative in the direction of 𝑥 − 𝑥 ,0
tive function with (noisy) observations of directional derivatives,6 which
evaluated at the midpoint (𝑥 + 𝑥 0 )/2.
are only weakly correlated with the corresponding function values.7
7 m. a. osborne et al. (2009). Gaussian Processes
Another option is to appeal to iterative numerical methods, discussed be- for Global Optimization. lion 3.
low, which actually benefit from having numerous eigenvalues clustered
around zero.

Iterative numerical methods

Direct methods can handle datasets of perhaps a few tens of thousands
of observations before the cubic scaling becomes too much to bear. In
most settings where Bayesian optimization would be considered, the

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 203

 implementation

cost of observation will preclude obtaining a dataset anywhere near this

size, in which case we need not consider the issue further. However,
when necessary, we may appeal to more complex approximate inference
schemes to scale to larger datasets.
One line of work in this direction is to solve the linear systems arising
in the posterior with iterative rather than direct numerical methods. The
method of conjugate gradients is especially well-suited as it is designed for
8 m. r. hestenes and e. stiefel (1952). Meth- symmetric positive-definite systems such as appear in the gp posterior.8
ods of Conjugate Gradients for Solving Linear The main idea behind the conjugate gradient method is to reinterpret
Systems. Journal of Research of the National
Bureau of Standards 49(6):409–436.
the solution of the linear system Cx = b as the solution of a related
and particularly well-behaved convex optimization problem.9 With this
9 If Cx = b, then |Cx − b | = 0. As this norm is
nonnegative and only vanishes at x, squaring
insight, we may derive a simple procedure to construct a sequence of
gives vectors {x𝑖 } guaranteed to converge in finite time to the desired solution.
For a system of size 𝑛, each iteration of this procedure requires only
x = arg min y>Cy − 2b>y.
y O(𝑛 2 ) work; the most expensive operation is a single matrix–vector
Thus x is also the solution of an unconstrained multiplication with C.
quadratic optimization problem, which is con- The method of conjugate gradients is guaranteed to converge (up to
vex as C is symmetric positive definite. round-off error) after 𝑛 iterations, but this does not offer any speedup over
direct methods. However, when C has a well-behaved spectrum – that is,
it is well-conditioned and/or has clustered eigenvalues – the sequence
converges rapidly. In many cases we may terminate after only 𝑘  𝑛
iterations with an accurate estimate of the solution, for an effectively
quadratic running time of O(𝑘𝑛 2 ). Although many covariance matrices
arising in practice are not necessarily well-conditioned, we may use a
technique known as preconditioning to transform a poorly conditioned
10 g. h. golub and c. f. van loan (2013). Matrix matrix to speed up convergence, with only minor overhead.10, 11
Computations. Johns Hopkins University Press. The use of conjugate gradients for gp inference can be traced back to
[§ 11.5]
the doctoral work of gibbs.12 Numerous authors have provided enhance-
11 k. cutajar et al. (2016). Preconditioning Ker- ments in the intervening years, and there is a now a substantial body
nel Matrices. icml 2016.
of related work. A good starting point is the work of gardner et al.,
12 m. n. gibbs (1997). Bayesian Gaussian Pro-
cesses for Regression and Classification. PhD
who provide a review of the literature and the key ideas from numerical
thesis. University of Cambridge. linear algebra required for large-scale gp inference.13 The authors refine
13 j. r. gardner et al. (2018). GPyTorch: Black-
these tools to exploit modern massively parallel hardware and build
box Matrix–Matrix Gaussian Process Infer- an accompanying software package scaling inference to hundreds of
ence with gpu Acceleration. neurips 2018. thousands of observations.

Sparse approximations
An alternative approach for scaling to large datasets is sparse approxi-
mation. Here rather than approximating the linear algebra arising in the
exact posterior, we approximate the posterior distribution itself with a
Gaussian process admitting tractable computation with direct numerical
methods. A large family of sparse approximations have been proposed,
which differ in their details but share the same general approach.
As we have seen, specifying an arbitrary Gaussian distribution for
a set of values jointly Gaussian distributed with a function of interest
Gaussian approximate inference: § 2.8, p. 301 induces a gp posterior consistent with that belief (2.39). This is a powerful
tool, used in approximate inference to optimize the fit of the induced

204
 9.1. gaussian process inference, scaling, and approximation

Gaussian process to a true, intractable posterior. Sparse approximation

methods make use of this property as well, but to achieve computational
rather than mathematical tractability. The idea is to craft a Gaussian
belief for a sufficiently small set of values such that the induced posterior
is a faithful, but tractable approximation to the true posterior.
For this discussion it is important that we explicitly account for
any observation noise that may be present in the values we wish to
condition on, as only independent noise – that is, with diagonal error
covariance – is suitable for sparse approximation. Consider conditioning
a Gaussian process on a vector of 𝑛 values z, for large 𝑛. We will assume observed values, z; |z | = 𝑛
the observation model z = y + 𝜺, where y is a vector of jointly Gaussian
distributed values as in (9.1), and 𝜺 is a vector of independent, zero-mean
Gaussian measurement noise with diagonal covariance matrix N. diagonal noise covariance, N
The first step of sparse approximation is to identify a set of 𝑚  𝑛
values 𝝊, called inducing values, whose distribution can in some sense inducing values, 𝝊; |𝝊 | = 𝑚  𝑛
capture most of the information in the full dataset. We will discuss
the selection of inducing values shortly; for the moment we assume an
arbitrary set has been chosen. The joint prior distribution of the observed
and inducing values is Gaussian:
     !
𝝊 𝝁 K>
; (9.3)
𝚺
𝑝 (𝝊, z) = N , ,
z m K C+N

and we will write the cross-covariance function for 𝝊 as

𝑘 (𝑥) = cov[𝝊, 𝜙 | 𝑥].

Conditioning (9.3) on z would yield the true Gaussian posterior on

the inducing values, but computing this posterior would be intractable. In
a sparse approximation, we instead prescribe a computationally tractable
posterior for 𝝊 informed by the available observations:
˜
˜ 𝚺).
𝑝 (𝝊 | z) ≈ 𝑞(𝝊 | z) = N (𝝊; 𝝁,

This assumed distribution then induces (hence the moniker inducing

values) a Gaussian process posterior on 𝑓 :
∫
𝑝 (𝑓 | z) ≈ 𝑝 (𝑓 | 𝝊) 𝑞(𝝊 | z) d𝝊 = GP (𝑓 ; 𝜇D , 𝐾D ),

which represents the sparse approximation. After transliteration of no-

tation, the posterior moments take the same form as (2.38), and the cost
of computation now scales according to the number of inducing values.
To complete this approximation scheme, we must specify a procedure 14 This is not strictly necessary. We could con-
for identifying a set of inducing values 𝝊 as well as the approximate sider using other values such as derivatives as
inducing values for added flexibility.
posterior 𝑞(𝝊 | z), and it is in these details that the various available
methods differ. The inducing values are usually taken to be function
values at a set of locations 𝝃 called pseudo- or inducing points, taking pseudopoints, inducing points
𝝊 = 𝑓 (𝝃 ).14 These inducing points are fictitious and do not need to
coincide with any actual observations. Once a suitable parameterization

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 205

 implementation

posterior mean, exact

posterior 95% credible interval, exact
Figure 9.1: Sparse approximation. Top:
the exact posterior belief
(about noisy observations
rather than the latent
function) for a Gaussian
process conditioned on
200 observations. Bottom:
a sparse approximation posterior mean, approximate
(9.5–9.7) using ten inducing posterior 95% credible interval, approximate
values corresponding to the
indicated inducing points,
designed to minimize the
kl divergence between the
induced and true posterior
distributions (9.4).

inducing points, 𝝃

of the inducing values is chosen, we usually design them – as well as

their inducing distribution – by optimizing a measure of fit between the
true posterior distribution and the resulting approximation.
titsias introduced a variational approach that has gained promi-
15 m. titsias (2009). Variational Learning of In- nence.15 The idea is to minimize the Kullback–Leibler divergence between
ducing Variables in Sparse Gaussian Processes. the true and induced posteriors on y and the inducing values 𝝊:
aistats 2009.
 
𝐷 kl 𝑞(y, 𝝊 | z) k 𝑝 (y, 𝝊 | z) , (9.4)
where
𝑞(y, 𝝊 | z) = 𝑝 (y | 𝝊) 𝑞(𝝊 | z)
is the implied joint distribution in the approximate posterior. Once opti-
mal inducing values are determined, the optimal inducing distribution is
as well, and examining the resulting approximate posterior gives insight
into the typical behavior of sparse approximations. The approximate
posterior mean is
 >
𝜇D (𝑥) = 𝜇 (𝑥) + K𝚺−1𝑘 (𝑥) (K𝚺−1 K> + N) −1 (z − m). (9.5)
Although this expression nominally entails solving a linear system of
size 𝑛, the low-rank-plus-diagonal structure of the matrix K𝚺−1 K> + N
16 For example, we may appeal to the Woodbury allows the system to be solved in time O(𝑛𝑚 2 ), merely linear in the size
identity and recognize that dealing with the of the dataset.16 With this favorable cost, sparse approximation can scale
diagonal matrix N is trivial. This is the reason Gaussian process inference to millions of observations without issue.
why we restricted the present discussion to
diagonal error covariance. Comparing this expression (9.5) with the true posterior mean:
𝜇D (𝑥) = 𝜇 (𝑥) + 𝜅 (𝑥) > (C + N) −1 (z − m),
we can identify a key approximation to the covariance structure of the
gp prior. Namely, we approximate the covariance function with:
𝐾 (𝑥, 𝑥 0) ≈ 𝑘 (𝑥) > 𝚺−1𝑘 (𝑥 0) = cov[𝑥, 𝝊] cov[𝝊, 𝝊] −1 cov[𝝊, 𝑥 0]. (9.6)

206
 9.2. optimizing acqisition functions

that is, we assume that all covariance between function values is mod-
erated through the inducing values. This is a popular approximation
scheme known as the Nyström method.17 Importantly, however, we note 17 c. k. i. williams and m. seeger (2000). Using
that the posterior mean does still reflect the information contained in the Nyström Method to Speed Up Kernel Ma-
chines. neurips 2000.
the entire dataset through the true residuals (z − m) and noise N. The
approximate posterior covariance also reflects this approximation:
 >  
𝐾 (𝑥, 𝑥 0) − K𝚺−1𝑘 (𝑥) (K𝚺−1 K> + N) −1 K𝚺−1𝑘 (𝑥 0) (9.7)
≈ 𝐾 (𝑥, 𝑥 0) − 𝜅 (𝑥) > (C + N) −1 𝜅 (𝑥 0).

Figure 9.1 illustrates a sparse approximation for a toy example fol-

lowing this approach. Here the inducing values were taken to be the
function17 values at ten inducing points, and both the inducing points and
their distribution were designed to minimize the kl divergence between
the true and induced posterior distributions (9.4). The approximation is
faithful: the posterior mean is nearly identical to the true mean and the
posterior credible intervals only display some minor differences from 19 a. j. smola and b. schölkopf (2000). Sparse
the truth. Increasing the number of inducing values would naturally Greedy Matrix Approximation for Machine
Learning. icml 2000.
improve the approximation.
Sparse approximation for Gaussian processes has a long history, 20 e. snelson and z. ghahramani (2005).
Sparse Gaussian Processes using Pseudo-
beginning in earnest with investigation into the Nyström approximation inputs. neurips 2005.
(9.6).18, 19 In addition to the variational approach mentioned above, an
21 j. hensman et al. (2015). mcmc for Variationally
approximation known as the fully independent training conditional (fitc) Sparse Gaussian Processes. neurips 2015.
approximation has also received significant attention20 and gives rise to
a similar approximate posterior. hensman et al. provided a variational
sparse approximation for non-Gaussian observation models, allowing
for scaling general gp latent models to large datasets.21

9.2 optimizing acqisition functions

In our discussion on computing optimization policies with Gaussian pro- computing common policies with Gaussian
cesses, we considered the pointwise evaluation of common acquisition processes: chapter 8, p. 157
functions and their gradient with respect to the proposed observation
location. However, we realize an optimization policy via the global opti-
mization of an acquisition function:

𝑥 ∈ arg max 𝛼 (𝑥 0; D). (9.8)

𝑥 0 ∈X

Every common Bayesian acquisition function is nonconvex in general,

so we must resort to some generic global optimization routine for this
inner optimization. Some care is required to guarantee success in this
optimization, as the behavior of a typical acquisition function can make
it a somewhat unusual objective function. In particular, consider a pro-
totypical Gaussian process model combining:
• a constant mean function (3.1),
• a stationary covariance function decaying to zero as |𝑥 − 𝑥 0 | → ∞, and stationarity: § 3.2, p. 50
• independent, homoskedastic observation noise (2.16).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 207

 implementation

For such a model, the prior predictive distribution 𝑝 (𝑦 | 𝑥) is identical

regardless of location. However, the posterior predictive distribution
𝑝 (𝑦 | 𝑥, D) also degenerates to the prior for locations sufficiently far
from observed locations, due to the decay of the covariance function.
In these regions, the gradients of the posterior predictive parameters
effectively vanish:
𝜕𝜇 𝜕𝑠
≈ 0; ≈ 0.
𝜕𝑥 𝜕𝑥
As a result, the gradient of the acquisition function (8.6) vanishes as well!
This vanishing gradient is especially problematic in high-dimensional
spaces, where the acquisition function will be flat on an overwhelming
fraction of the domain unless the prior encodes absurdly long-scale
correlations, a consequence of the unshakable curse of dimensionality.
curse of dimensionality: § 3.5, p. 61 Thus, the acquisition function will only exhibit interesting behavior
in the neighborhood of previous observations – where the posterior
predictive distribution is nontrivial – and it is here we should spend
most of our effort during optimization.

Optimization approaches
There are two common lines of attack for optimizing acquisition func-
tions in Bayesian optimization. One approach is to use an off-the-shelf
derivative-free global optimization method such as the “dividing rectan-
22 d. r. jones et al. (1993). Lipschitzian Optimiza- gles” (direct) algorithm of jones et al.22 or a member of the covariance
tion Without the Lipschitz Constant. Journal matrix adaptation evolution strategy (cma–es) family of algorithms.23
of Optimization Theory and Application 79(1):
157–181.
Although a popular choice in the literature, we argue that neither is a
particularly good choice in situations where the acquisition function may
23 n. hansen (2016). The cma Evolution Strategy:
A Tutorial. arXiv: 1604.00772 [cs.LG].
devolve into effective flatness as described above. However, an algorithm
of this class may be reasonable in modest dimension where optimization
can be somewhat exhaustive.
In general, a better alternative is multistart local optimization, mak-
ing use of the gradients computed in the previous chapter for rapid
convergence. kim and choi provided an extensive comparison of the
theoretical and empirical performance of global and local optimization
routines for acquisition function optimization, and ultimately recom-
24 j. kim and s. choi (2020). On Local Optimiz- mended multistart local optimization as best practice.24 This is also the
ers of Acquisition Functions in Bayesian Opti- approach used in (at least some) sophisticated modern software packages
mization. 12458:675–690.
for Bayesian optimization.25
25 m. balandat et al. (2020). BoTorch: A Frame- To ensure success, we must carefully select starting points to ensure
work for Efficient Monte-Carlo Bayesian Op-
timization. neurips 2020.
that the relevant regions of the domain are searched. jones (the same
jones of the direct algorithm) recognized the problem of vanishing
gradients described above and suggested a simple heuristic for select-
ing local optimization starting points by enumerating and pruning the
26 d. r. jones (2001). A Taxonomy of Global Op- midpoints between all pairs of observed points.26 A more brute-force
timization Methods Based on Response Sur- approach is to measure the acquisition function on a exhaustive covering
faces. Journal of Global Optimization 21(4):345–
383.
of the domain – generated for example by a low-discrepancy sequence
– then begin local searches from the highest values seen. This can be
effective if the initial set of points is dense enough to probe the neigh-

208
 9.2. optimizing acqisition functions

borhoods of previous observations; otherwise, it would be prudent to

augment with a locally motivated approach such as jones’s.
Multistart local optimization has the advantage of being embarrass- 27 e. brochu et al. (2010). A Tutorial on Bayes-
ingly parallel. Further, both global and multistart local optimization ian Optimization of Expensive Cost Functions,
with Application to Active User Modeling and
of the acquisition function can be treated as anytime algorithms that Hierarchical Reinforcement Learning. arXiv:
constantly improve their proposed observations until we are ready to 1012.2599 [cs.LG].
act.27

Optimization in latent spaces

A common approach for modeling on high-dimensional domains is modeling functions on high-dimensional
to apply some mapping from the domain to some lower-dimensional domains: § 3.5, p. 61
representation space, then construct a Gaussian process on that space.
With such a model, it is tempting to optimize an acquisition function 28 Detailed advice for this setting is provided by:
on the latent space rather than on the original domain so as to at least
z. wang et al. (2016b). Bayesian Optimization
partially sidestep the curse of dimensionality. This can be an effective in a Billion Dimensions via Random Embed-
approach when we can faithfully “decode” from the latent space back into dings. Journal of Artificial Intelligence Research
the original domain for evaluation, a process that can require careful 55:361–387.
thought even when the latent embedding is linear (3.30) due to the
nonbijective nature of the map.28 linear embeddings: § 3.5, p. 62
Fortunately, modern neural embedding techniques such as (varia- neural embeddings: § 3.5, p. 61, § 8.11, p. 199
tional) autoencoders provide a decoding mechanism as a natural side
effect of their construction. When we are fortunate enough to have suffi-
cient unlabeled data to learn a useful unsupervised representation prior
to optimization, we may simply optimize in the latent space and feed each
chosen observation location through the decoder. gómez-bombarelli
et al. for example applied Bayesian optimization to de novo molecular
design.29 Their model combined a pretrained variational autoencoder for de novo molecular design: appendix d, p. 314
molecular structures with a Gaussian process on the latent embedding
space; the autoencoder was trained on a large precompiled database of 29 r. gómez-bombarelli et al. (2018). Automatic
known molecules.30 A welcome side effect of this construction was that, Chemical Design Using a Data-Driven Contin-
uous Representation of Molecules. acs Central
by optimizing the acquisition function over the continuous embedding Science 4(2):268–276.
space, the decoding process had the freedom to generate novel structures
30 j. j. irwin et al. (2020). zinc20 – A Free
not seen in the autoencoder’s training data. Ultralarge-Scale Chemical Database for Lig-
We may also pursue this approach even when we begin optimization and Discovery. Journal of Chemical Informa-
without any data at all. For example, moriconi et al. demonstrated tion and Modeling 60(12):6065–6073.
success jointly learning a nonlinear low-dimensional representation as 31 r. moriconi et al. (2020). High-dimensional
well as a corresponding decoding stage throughout optimization on the Bayesian optimization using low-dimensional
feature spaces. Machine Learning 109(9–10):
fly.31
1925–1943.

Optimization on combinatorial domains

Combinatorial domains present a challenge for Bayesian optimization as
the optimization policy (9.8) requires combinatorial optimization of the
acquisition function. It is difficult to provide concrete advice for such
a situation, as the details of the domain may in some cases suggest a
natural path forward. One potential solution is outlined above: when
the domain is a space of discrete structured objects such as graphs (say,

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 209

 implementation

molecules), we might find a useful continuous embedding of the domain

and simply work there instead.29
gene and protein design: appendix d, p. 319 Applications such as gene or protein design entail combinatorial
optimization over strings, whose nature presents problems both in mod-
eling and optimizing acquisition functions, although optimizing a neural
32 b. l. hie and k. k. yang (2021). Adaptive ma- encoding remains a viable option.32 moss et al. described a more direct
chine learning for protein engineering. arXiv: approach for Bayesian optimization over strings when this is not possi-
2106.05466 [q-bio.QM].
ble.33 The objective function was modeled by a Gaussian process with a
33 h. b. moss et al. (2020). boss: Bayesian Opti- special-purpose string kernel, and the induced acquisition function was
mization over String Spaces. neurips 2020.
then optimized using genetic algorithms, which are well-suited for string
optimization. The authors also showed how to incorporate a context-free
grammar into the optimization policy to impose desired constraints on
the structure of the strings investigated.
In the general case, we note that our previous sketch of how a “typical”
acquisition function behaves with a “typical” model – with its most
interesting behavior near observed data – carries over to combinatorial
domains and may lead to useful heuristics. For example, rather than
34 As evaluated by the covariance function. enumerating the domain (presumably impossible) or sampling from the
35 r. garnett et al. (2010). Bayesian Optimiza- domain (presumably yielding poor coverage), we might instead curate
tion for Sensor Set Selection. ipsn 2010. a small list of candidate points offering options for both exploitation
36 r. baptista and m. poloczek (2018). Bayes- and exploration. We could perhaps generate a list of points similar to34
ian Optimization of Combinatorial Structures. the thus-far best-seen points, encouraging exploitation, and augment
with a small set of points constructed to cover the domain, encouraging
icml 2018.
37 c. oh et al. (2019). Combinatorial Bayesian Op-
timization using the Graph Cartesian Product.
exploration. This approach was used for example by garnett et al. for
neurips 2019. Bayesian set function optimization.35
38 j. kim et al. (2021). Bayesian optimization with
Other approaches are also possible. For example, several authors have
approximate set kernels. Machine Learning constructed gp models for particular combinatorial spaces whose struc-
110(5):857–879. ture simplifies the (usually approximate but near-optimal) optimization
of particular acquisition functions induced from the model.36, 37, 38

The benefit of imperfect optimization?

We conclude this discussion with one paradoxical remark. In some cases
it may actually be advantageous to optimize an acquisition function
imperfectly. As many common policies are based on extremely myopic
(one-step) reasoning, imperfect optimization may yield a useful explo-
ration benefit as a side effect. This phenomenon, and the effect it may
have on empirical performance, deserves thoughtful consideration.

9.3 starting and stopping optimization

Finally, we briefly consider the part of optimization that happens outside
the application of an optimization policy: initialization and termination.

Initialization
Theoretically, one can begin a Bayesian optimization routine with a
completely empty dataset D = ∅ and then use an optimization policy to

210
 9.3. starting and stopping optimization

design every observation, and indeed this has been our working model
of Bayesian optimization since sketching the basic idea in algorithm 1.1.
However, Bayesian optimization policies are informed by an underlying
belief about the objective function, which can be significantly misin-
formed when too little data is available, especially when relying on point
estimation for model selection rather than accounting for (significant!) model averaging: § 4.4, p. 74
uncertainty in model hyperparameters and/or model structures.
Due to the sequential nature of optimization and the dependence of 39 These approaches are discussed and evaluated
each decision on the data observed in previous iterations, it can be wise in:
to use a model-independent procedure to design a small number of initial m. w. hoffman and b. shahriari (2014). Mod-
observations before beginning optimization in earnest. This procedure ular mechanisms for Bayesian optimization.
can be as simple as random sampling39 or a space-filling design such as a Bayesian Optimization Workshop, neurips 2014.
low-discrepancy sequence or Latin hypercube design.40 When repeatedly 40 d. r. jones et al. (1998). Efficient Global Opti-
solving related optimization problems, we may even be able to learn how mization of Expensive Black-Box Functions.
Journal of Global Optimization 13(4):455–492.
to initialize Bayesian optimization routines from experience. Some au-
thors have proposed sophisticated “warm start” initialization procedures
for hyperparameter tuning using so-called metafeatures characterizing
the datasets under consideration.41 41 m. feurer et al. (2015). Initializing Bayes-
ian Hyperparameter Optimization via Meta-
Learning. aaai 2015.
Termination
In many applications of Bayesian optimization, we assume a preallocated
budget on the number of observations we will make and simply terminate
optimization when that budget is expended. However, we may also treat
termination as a decision and adaptively determine when to stop based optimal stopping rules: § 5.4, p. 103
on collected data. Of course, in practice we are free to design a stopping
rule however we see fit, but we can outline some possible options.
Especially when using a policy grounded in decision theory, it is
natural to terminate optimization when the maximum of our chosen
acquisition function drops below some threshold 𝑐, which may depend
on 𝑥:
 
 
max 𝛼 (𝑥; D) − 𝑐 (𝑥) < 0 . (9.9)
𝑥 ∈X

For acquisition functions derived from decision theory, such a stopping 42 See in particular our discussion of the one-step
rule may be justified theoretically: we stop when the expected gain from optimal stopping rule, p. 104.
the optimal observation is no longer worth the cost of acquisition.42 A 43 Some early (but surely not the earliest!) exam-
majority of the stopping rules described in the literature assume this ples:
form, with the threshold 𝑐 often being determined dynamically based d. d. cox and s. john (1992). A Statistical
on the scale of observed data.43 dai et al. combined a stopping rule of Method for Global Optimization. smc 1992.
this form with an otherwise non-decision-theoretic policy (gp-ucb) and
d. r. jones et al. (1998). Efficient Global Opti-
showed that its asymptotic performance in terms of expected regret was mization of Expensive Black-Box Functions.
not adversely affected despite the mismatch in motivation between the Journal of Global Optimization 13(4):455–492.
policy and stopping rule.44 44 z. dai et al. (2019). Bayesian Optimization
It may also be prudent to consider purely data-dependent stopping Meets Bayesian Optimal Stopping. icml 2019.
rules in order to avoid undue expense arising from miscalibrated models 45 l. acerbi and w. j. ma (2017). Practical Bayes-
fruitlessly continuing optimization based on incorrect beliefs. For exam- ian Optimization for Model Fitting with Bayes-
ple, acerbi and ma proposed augmenting a bound on the total number of ian Adaptive Direct Search. neurips 2017.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 211

implementation

observations with an early stopping option if no optimization progress

is made over a given number of observations.45

9.4 summary of major ideas

• Gaussian process inference requires solving a system of linear equations
whose size grows with the number of observed values (9.2).
direct computation via Cholesky • A direct implementation via the Cholesky decomposition is a straightfor-
decomposition: § 9.1, p. 201 ward option, but scales cublically with the number of observed values.
• Scaling to larger datasets is possible by appealing to iterative numerical
iterative numerical methods: § 9.1, p. 203 methods such as the method of conjugate gradients, or to sparse approxi-
sparse approximation: § 9.1, p. 204 mation, where we approximate an intractable posterior Gaussian process
with a Gaussian process conditioned on carefully designed, fictitious
observations at locations called inducing points. These methods can scale
inference to hundreds of thousands of observations or more.
the potential for vanishing gradients: § 9.2, • When optimizing acquisition functions, it is important to be aware of the
p. 207 potential for vanishing gradients in extrapolatory regions of the domain
and plan accordingly.
initialization: § 9.3, p. 210 • It is usually a good idea to begin optimization with a small set of obser-
vations designed in a model-agnostic fashion in order to begin with a
somewhat informed model of the objective function.
termination: § 9.3, p. 211 • When dynamic termination is desired, simple schemes based on thresh-
olding the acquisition function can be effective.

212
 10
THEORETICAL ANALYSIS

The Bayesian optimization procedures we have developed throughout

this book have demonstrated remarkable empirical performance in a
huge swath of practical settings, many of which are outlined in appendix
d. However, good empirical performance may not be enough to satisfy
those who value rigor over results. Fortunately, many Bayesian opti- annotated bibliography of applications:
mization algorithms are also backed by strong theoretical guarantees appendix d, p. 313
on their performance. The literature on this topic is now quite vast, and
convergence has been studied by different authors in different ways,
sometimes involving slight nuances in approach and interpretation. We
will take an in-depth look at this topic in this chapter, covering the most
common lines of attack and outlining the state-of-the-art in results.
A running theme throughout this chapter will be understanding how
various measures of optimization error decrease asymptotically as an
optimization policy is repeatedly executed. To facilitate this discussion,
we will universally use 𝜏 in this chapter to indicate dataset size, or size of dataset, 𝜏
equivalently to indicate the number of steps an optimization procedure
is assumed to have run. As we will primarily be interested in asymptotic
results as 𝜏 → ∞, this convention does not rule out the possibility
of starting optimization with some arbitrary dataset of fixed size. We
will use subscripts to indicate dataset size when necessary, notating the
dataset comprising the first 𝜏 observations with: dataset after 𝜏 observations, D𝜏
 𝜏
D𝜏 = (x𝜏 , y𝜏 ) = (𝑥𝑖 , 𝑦𝑖 ) 𝑖=1 .

When studying the convergence of a global optimization algorithm, what does it mean to converge?
we must be careful to define exactly what we mean by “convergence.” In
general, a convergence argument entails:
• choosing some measure of optimization error,
• choosing some space of possible objective functions, and
• establishing some guarantee for the chosen error on the chosen function
space, such as an asymptotic bound on the worst- or average-case error
in the large-sample limit 𝜏 → ∞.
There is a great deal of freedom in the last of these steps, and we will
discuss several important results and proof strategies later in this chapter.
However, there are well-established conventions for the first two of
these steps, which we will introduce in the following two sections. We regret: below
begin with the notion of regret, which provides a natural measure of useful spaces of objective functions: § 10.2,
optimization error. p. 215

10.1 regret
Regret is a core concept in the analysis of optimization algorithms, Bayes-
ian or otherwise. The role of regret is to quantify optimization progress
in a manner suitable for establishing convergence to the global optimum
and studying the rate of this convergence. There are several definitions

This material will be published by Cambridge University Press as Bayesian Optimization. 213
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 theoretical analysis

1 One might propose an alternative definition of regret used in different contexts, all based on the same idea: comparing
of error by measuring how closely the ob- the objective function values visited during optimization to the globally
served locations approach a global maximum
𝑥 ∗ rather than by how closely the observed optimal value, 𝑓 ∗. The larger this gap, the more “regret” we incur in
values approach the value of the global opti- retrospect for having invested in observations at suboptimal locations.1
mum 𝑓 .∗ However, it turns out this is both less Regret is an unavoidable consequence of decision making under un-
convenient for analysis and harder to moti- certainty. Without foreknowledge of the global optimum, we must of
vate: in practice it is the value of the objective
that we care about the most, and discovering course spend some time searching for it, and even what may be optimal
a near-optimal value is a success regardless of actions in the face of uncertainty may seem disappointing in retrospect.2
its distance to the global optimum. However, such actions are necessary in order to learn about the environ-
2 As outlined in chapter 5 (p. 87), optimal actions ment and inform future decisions. This reasoning gives rise to the classic
maximize expected utility in the face of uncer- tension between exploration and exploitation in policy design: although
tainty. Many observations may not result in
exploration may not yield immediate progress, it enables future success,
progress, even though designed with the best
of intentions. and if we are careful, reduces future regret. Of course, exploration alone
3 For example, it is easy to develop a space- is not sufficient to realize a compelling optimization strategy,3 as we must
filling design that will eventually locate the also exploit what we have learned and adapt our behavior accordingly.
global optimum of any continuous function An ideal algorithm thus explores efficiently enough that its regret can at
through sheer dumb luck – but it might not
least be limited in the long run.
do so very quickly!
Most analysis is performed in terms of one of two closely related
notions of regret: simple or cumulative regret, defined below.

Simple regret
4 The measures of regret introduced here do Let D𝜏 represent some set of (potentially noisy) observations gathered
not depend on the observed values y but only during optimization,4 and let 𝝓𝜏 = 𝑓 (x𝜏 ) represent the objective function
on the underlying objective values 𝝓. We are
making a tacit assumption that y is sufficiently
values at the observed locations. The simple regret associated with this
informative about 𝝓 for this to be sensible in data is the difference between the global maximum of the objective and
the large-sample limit. the maximum restricted to the observed locations:5
5 Occasionally a slightly different definition of
simple regret is used, analogous to the global 𝑟𝜏 = 𝑓 ∗ − max 𝝓𝜏 . (10.1)
reward (6.5), where we measure regret with
respect to the maximum of the posterior mean:
It is immediate from its definition that simple regret is nonnegative
𝑟𝜏 = 𝑓 ∗ − max 𝜇D𝜏 (𝑥). and vanishes only if the data contain a global optimum. With this in
𝑥 ∈X
mind, a common goal is to show that the simple regret of data obtained
by some policy approaches zero, implying the policy will eventually (and
convergence goal: show 𝑟𝜏 → 0 perhaps efficiently) identify the global optimum, up to vanishing error.

Cumulative regret
To define cumulative regret, we first introduce the instantaneous regret 𝜌
corresponding to an observation at some point 𝑥, which is the difference
instantaneous regret, 𝜌 between the global maximum of the objective and the function value 𝜙:

𝜌 = 𝑓 ∗ − 𝜙. (10.2)

cumulative regret of D𝜏 , 𝑅𝜏 The cumulative regret for a dataset D𝜏 is then the total instantaneous
regret incurred:
∑︁ ∑︁
𝑅𝜏 = 𝜌𝑖 = 𝜏𝑓 ∗ − 𝜙𝑖 . (10.3)
𝑖 𝑖

214
 10.2. useful function spaces for studying convergence

Relationship between simple and cumulative regret

Simple and cumulative regret are analogous to the simple (6.3) and 6 Note that neither notion of regret represents a
cumulative (6.7) reward utility functions, and any intuition regarding valid utility function itself, as in general both
𝑓 ∗ and 𝝓 would be random variables in that
these utilities transfers to their regret counterparts.6 context.
However, these two definitions of regret are not directly comparable, 7 For a deeper discussion on the connections
even on the same data – for starters, simple regret is nonincreasing as between simple and cumulative regret, see:
more data is collected, whereas cumulative regret is nondecreasing. How-
s. bubeck et al. (2009). Pure Exploration in
ever, suitable normalization allows some useful comparison.7 Namely, Multi-armed Bandits Problems. alt 2009.
consider the average, rather than cumulative, regret:
𝑅𝜏 1 ∑︁
= 𝑓∗ − 𝜙𝑖 .
𝜏 𝜏 𝑖

As the mean of a vector is a lower bound on its maximum, we may derive

an upper bound on simple regret in terms of cumulative regret:
𝑅𝜏
𝑟𝜏 ≤ . (10.4)
𝜏
In this light, a common goal is to show that an optimization algorithm
has the so-called no-regret property, which means that its average regret convergence goal: show 𝑅𝜏 /𝜏 → 0
vanishes with increasing data:
𝑅𝜏
lim = 0. (10.5)
𝜏→∞ 𝜏
Equivalently, a policy achieves no regret if its cumulative regret grows no-regret property implies convergence in
sublinearly with the dataset size 𝜏. This is sufficient to prove convergence simple regret
in terms of simple regret as well by appealing to the squeeze theorem:
𝑅𝜏
0 ≤ 𝑟𝜏 ≤ → 0.
𝜏
Although no regret guarantees convergence in simple regret, the convergence in simple regret does not imply
reverse is not necessarily the case. It is easy to find counterexamples the no-regret property
– consider, for example, modifying a no-regret policy to select a fixed
suboptimal point every other observation. The simple regret would still
vanish (half as fast), but constant instantaneous regret on alternating
iterations would prevent sublinear cumulative regret. Thus the no-regret
property is somewhat stronger than convergence in simple regret alone.
From another perspective, simple regret is more tolerant of exploration:
we only need to visit the global optimum once to converge in terms
of simple regret, whereas effectively all observations must eventually
become effectively optimal to achieve the no-regret property.

10.2 useful function spaces for studying convergence

Identifying the “right” function space to consider when studying conver-
gence is a subtle decision, as we must strike a balance between generality
and practicality. One obvious choice would be the space of all continuous

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 215

 theoretical analysis

𝑓 𝑔 𝑓 +𝑔

𝐵 𝐵 𝐵

Figure 10.1: Modifying a continuous function 𝑓 (left) to feature a “needle” on some ball (here an interval) 𝐵. We construct
a continuous function 𝑔 vanishing on the compliment of 𝐵 (middle), then add this “correction” to 𝑓 (right).

functions. However, it turns out this space is far too large to be of much
theoretical interest, as it contains functions that are arbitrarily hard to
optimize. Nonetheless, it is easy to characterize convergence (in terms
of simple regret) on this space, and some convergence guarantees of this
type are known for select Bayesian optimization algorithms. We will
begin our discussion here.
convergence on “nice” continuous functions: We may gain some traction by considering a family of more plausible,
p. 218 “nice” functions whose complexity can be controlled enough to guarantee
rapid convergence. In particular, choosing a Gaussian process prior
for the objective function implies strong correlation structure, and this
insight leads to natural function spaces to study. This has become the
standard approach in modern analysis, and we will consider it shortly.

Convergence on all continuous functions

The largest reasonable space we might want to consider is the space of
all continuous functions. However, continuous functions can be poorly
behaved from the point of view of optimization, and we cannot hope
for strong convergence guarantees as a result. There is simply too much
freedom for functions to “hide” their optima in inconvenient places.
To begin, we establish a simple characterization of convergence on
all continuous functions in terms of eventual density of observation.
Theorem. Let X be a compact metric space. An optimization policy con-
verges in terms of simple regret on all continuous functions 𝑓 : X → ℝ if
Ð∞
and only if the set of eventually observed points 𝑖=1 {𝑥𝑖 } is always dense.
The proof is instructive as it shows what can go wrong with general
continuous functions. First, if the set of eventually observed points is
dense in X, we may construct a sequence of observations converging to a
8 For example by taking 𝑥𝑖 within a ball of ra-
dius 1/𝑖 around 𝑥 ∗. global maximum 𝑥 ∗. 8 The associated function values {𝜙𝑖 } then converge
to 𝑓 ∗ by continuity, and thus 𝑟𝜏 → 0.
9 Some care would be required to make the fol-
lowing argument rigorous for stochastic ob- If density fails9 for some continuous function 𝑓 – and thus there is
servations and/or policies, but the spirit would some ball 𝐵 ⊂ X that will never contain an observation – then we may
remain intact. foil the policy with a “needle in a haystack.” We construct a continuous

216
 10.2. useful function spaces for studying convergence

function 𝑔 vanishing on the complement of 𝐵 and achieving arbitrarily

high values on 𝐵.10 Adding the needle to 𝑓 creates a continuous function 10 For example, we may scale the distance to the
𝑓 + 𝑔 with arbitrary – and, once the needle is sufficiently tall, never complement of 𝐵.
observed – maximum. As 𝑓 and 𝑓 + 𝑔 agree outside 𝐵, the policy cannot
distinguish between these two functions, so we can find continuous
functions for which the policy has arbitrarily high simple regret.
We can use this strategy of building adversarial “needles in haystacks” worst-case simple regret unbounded
to further show that, even if we can guarantee convergence on all con-
tinuous functions, we cannot hope to demonstrate rapid convergence in
simple regret, at least not in the worst case. Unless the domain is finite,
running any policy for any finite number of iterations will leave unob-
served “holes” that we can fill in with arbitrarily tall “needles,” and thus
the worst-case regret will be unbounded at every stage of the algorithm.

Convergence results for all continuous functions on the unit interval

Establishing the density criterion above has proven difficult for Bayesian
optimization algorithms in general spaces with arbitrary models. How-
ever, restricting the domain to the unit interval X = [0, 1] and limiting
the objective function and observation models to certain well-behaved
combinations has yielded universal convergence guarantees for some
Bayesian procedures.
For example, kushner sketched a proof of convergence in simple
regret when maximizing probability of improvement on the unit interval,
when the objective function model is the Wiener process and observa-
tions are exact or corrupted by additive Gaussian noise.11 žilinskas later 11 h. j. kushner (1964). A New Method of Locat-
provided a proof in the noiseless case.12 This particular model exhibits ing the Maximum Point of an Arbitrary Multi-
peak Curve in the Presence of Noise. Journal
a Markov property that enables relatively straightforward analysis by of Basic Engineering 86(1):97–106.
characterizing its behavior on each of the subintervals subdivided by the
12 a. g. žilinskas (1975). Single-Step Bayesian
data. This structure enables a simple proof strategy by assuming some Search Method for an Extremum of Functions
subinterval is never subdivided and arriving at a contradiction. of a Single Variable. Kibernetika (Cybernetics)
Convergence guarantees for this policy (under the name the “p- 11(1):160–166.
algorithm”) on the unit interval have also been established for smoother 13 j. calvin and a. žilinskas (1999). On the
models of the objective function – that is, with differentiable rather Convergence of the P-Algorithm for One-
Dimensional Glboal Optimization of Smooth
than merely continuous sample paths13 – including the once-integrated
Functions. Journal of Optimization Theory and
Wiener process.14 Convergence rates for these algorithms for general Applications 102(3):479–495.
continuous functions have also been derived.13, 14, 15 In light of the above
14 j. m. calvin and a. žilinskas (2001). On Con-
discussion, these convergence rates are not in terms of simple regret but vergence of a P-Algorithm Based on a Statisti-
rather in terms of shrinkage in the size of the subinterval containing cal Model of Continuously Differentiable Func-
the global optimum, and thus the distance from the closest observed tions. Journal of Global Optimization 19(3):229–
245.
location to the global optimum. The proofs of these results all relied on
exact observation of the objective function. 15 j. m. calvin (2000). Convergence Rate of the
P-Algorithm for Optimization of Continuous
Convergence on all continuous functions on the unit interval has also Functions. In: Approximation and Complex-
been established for maximizing expected improvement, in the special ity in Numerical Optimization: Continuous and
case of exact observation and a Wiener process model on the objective.16 Discrete Problems.
The proof strategy again relied on the special structure of the posterior 16 m. locatelli (1997). Bayesian Algorithms for
to show that the observed locations will always subdivide the domain One-Dimensional Global Optimization. Jour-
nal of Global Optimization 10(1):57–76.
such that no “hole” is left behind.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 217

 theoretical analysis

Convergence on “nice” continuous functions

In the pursuit of stronger convergence results, we may abandon the
space of all continuous functions and focus on some space of suitably
well-behaved objectives. By limiting the complexity of the functions
we consider, we can avoid complications arising from adversarial exam-
ples like “needles in haystacks.” This can be motivated from the basic
assumption that optimization is to be feasible at all, which is not the
17 All of the algorithms we will study for the case when facing an adversary creating arbitrarily difficult problems.
remainder of the chapter will use a Gaussian Instead, we may seek strong performance guarantees when optimizing
process belief on the objective function, where
the theory is mature. “plausible” objective functions. Conveniently, a Gaussian process model
on the objective GP (𝑓 ; 𝜇, 𝐾) gives rise to several paths forward.17

Sample paths of a Gaussian process

sample path continuity: § 2.5, p. 28 In a Bayesian analysis, it is natural to assume that the objective function
is a sample path from the Gaussian process used to model it. Assuming
sample path continuity, sample paths of a Gaussian process are much bet-
ter behaved than general continuous functions. The covariance function
provides regularization on the behavior of sample paths via the induced
correlations among function values; see the figure in the margin. As a
result, we can ensure that functions with exceptionally bad behavior
Sample paths of a stationary Gaussian process (such as hidden “needles”) are also exceptionally rare.
with Matérn 𝜈 = 5/2 covariance (3.14) show The sample path assumption provides a known distribution for the
more regularity than arbitrary continuous function values at observed locations (2.2–2.3), allowing us to derive
functions.
average-case results. An important concept here is the Bayesian (or
simply Bayes) regret of a policy, which is the expected value of the
(simple or cumulative) regret incurred when following the policy, say
Bayesian (Bayes) regret for 𝜏 steps:
𝔼[𝑟𝜏 ]; 𝔼[𝑅𝜏 ]. (10.6)

18 In some analyses, we may also seek bounds This expectation is taken with respect to uncertainty in the objective
on the regret that hold with high probability
function 𝑓, the observation locations x, and the observed values y.18
with respect to these random variables rather
than bounds on the expected regret.
The worst-case alternative
The gp sample path assumption is not always desirable, for example in
the context of a frequentist (that is, worst-case) analysis of a Bayesian
optimization algorithm. This is not as contradictory as it may seem, as
Bayesian analyses can lack robustness to model misspecification – a cer-
tainty in practice. An alternative is to assume that the objective function
lies in some explicit space of “nice” functions H, then find worst-case
convergence guarantees for the Bayesian algorithm on inputs satisfying
this regularity assumption. For example, we might seek to bound the
worst-case regret on H after 𝜏 steps: 𝑟¯ (𝜏, H), worst-case expected (simple or cumulative) regret for a function in this
𝑅¯ (𝜏, H) space after 𝜏 decisions:

𝑟¯𝜏 [H] = sup 𝔼[𝑟𝜏 ]; 𝑅¯𝜏 [H] = sup 𝔼[𝑅𝜏 ]. (10.7)

𝑓 ∈H 𝑓 ∈H

218
 10.2. useful function spaces for studying convergence

Figure 10.2: Left: functions in the rkhs

corresponding to the Matérn
𝜈 = 1/2 covariance function
(3.11). Right: sample paths
from a gp with the same co-
variance.
Here the expectation is over with respect to the observed locations x
and the observed values y,19 but uncertainty in the objective function – 19 In the special case of exact observation and
presumably the most troubling factor in a Bayesian analysis – is replaced a deterministic policy, such a bound would
entail no probabilistic elements.
by a pessimistic bound on the functions in H. Note that the algorithms we
will analyze still use a Gaussian process belief on 𝑓 in making decisions,
but the objective function is not generated according to that model.

Reproducing kernel Hilbert spaces

Corresponding to every covariance function 𝐾 is a natural companion reproducing kernel Hilbert space
function space, its reproducing kernel Hilbert space (rkhs) H𝐾 . There is a corresponding to 𝐾, H𝐾
strong connection between this space and a centered Gaussian process
with the same covariance function, GP (𝑓 ; 𝜇 ≡ 0, 𝐾). Namely, consider
the set of functions of the form
∑︁
𝑛
𝑥 ↦→ 𝛼𝑖 𝐾 (𝑥𝑖 , 𝑥), (10.8)
𝑖=1

where {𝑥𝑖 } ⊂ X is an arbitrary finite set of input locations with cor- 20 Equivalently, this is the span of the set of co-
responding real-valued weights {𝛼𝑖 }.20 Note this is precisely the set of variance
 functions with one input held fixed:
span 𝑥 ↦→ 𝐾 (𝑥, 𝑥 0 ) | 𝑥 0 ∈ X .
all possible posterior mean functions for the Gaussian process arising
from exact inference!21 The rkhs H𝐾 is then the completion of this space 21 Inspection of the general posterior mean func-
tions in (2.14, 2.19) reveals they can always be
endowed with the inner product
(and can only be) written in this form.
∑︁
𝑛 ∑︁
𝑚  ∑︁
𝑛 ∑︁
𝑚
𝛼𝑖 𝐾 (𝑥𝑖 , 𝑥), 0
𝛽𝑗 𝐾 (𝑥 𝑗 , 𝑥) = 𝛼𝑖 𝛽𝑗 𝐾 (𝑥𝑖 , 𝑥 𝑗0). (10.9)
𝑖=1 𝑗=1 𝑖=1 𝑗=1

That is, the rkhs is roughly the set of functions “as smooth as” a pos-
terior mean function of the corresponding gp, according to a notion of
“explainability” by the covariance function.
It turns out that belonging to the rkhs H𝐾 is a stronger regularity
assumption than being a sample path of the corresponding gp. In fact,
unless the rkhs is finite-dimensional (which is not normally the case), 22 m. n. lukić and j. h. beder (2001). Stochas-
sample paths from a Gaussian process almost surely do not lie in the tic Processes with Sample Paths in Reproduc-
ing Kernel Hilbert Spaces. Transactions of the
corresponding rkhs:22, 23 Pr(𝑓 ∈ H𝐾 ) = 0. However, the posterior mean American Mathematical Society 353(10):3945–
function of the same process always lies in the rkhs by the above con- 3969.
struction. Figure 10.2 illustrates a striking example of this phenomenon: 23 However, sample paths do lie in a “slightly
sample paths from a stationary gp with Matérn covariance function larger” rkhs we can determine from the co-
with 𝜈 = 1/2 (3.11) are nowhere differentiable, whereas members of the variance function:
corresponding rkhs are almost everywhere differentiable. Effectively, m. kanagawa et al. (2018). Gaussian Processes
the process of averaging over sample paths “smooths out” their erratic and Kernel Methods: A Review on Connec-
behavior in the posterior mean, and elements of the rkhs exhibit similar tions and Equivalences. arXiv: 1807 . 02582
smoothness. [stat.ML] [theorem 4.12]

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 219

 theoretical analysis

Reproducing kernel Hilbert space norm

rkhs norm, k𝑓 k H𝐾 Associated with a rkhs H𝐾 is a norm k 𝑓 k H𝐾 that can be interpreted as
a measure of function complexity with respect to the covariance func-
tion 𝐾. The rkhs norm derives from the pre-completion inner product
(10.9), and we can build intuition for the norm by drawing a connection
between that inner product and familiar concepts from Gaussian process
regression. The key is the characterization of the pre-completion func-
tion space (10.8) as the space of all possible posterior mean functions for
the corresponding centered Gaussian process GP (𝑓 ; 𝜇 ≡ 0, 𝐾).
rkhs norm of posterior mean function To be more explicit, consider the posterior mean after observing a
dataset of exact observations D = (x, y), inducing the posterior mean
(2.14):
𝜇D (𝑥) = 𝐾 (𝑥, x)𝚺−1 y,
24 For this example we have 𝜶 = 𝚺−1 y in (10.8). where 𝚺 = 𝐾 (x, x). Then the (squared) rkhs norm of the posterior mean
is:24
k𝜇D k 2H𝐾 = h𝜇D , 𝜇D i = y> 𝚺−1 y. (10.10)
That is, the rkhs norm of the posterior mean is the Mahalanobis norm
of the observed data y under their Gaussian prior distribution (2.2–2.3).
connection between rkhs norm and gp We have actually seen this score before: it appears in the log marginal
marginal likelihood likelihood of the data under the Gaussian process (4.8), where we inter-
preted it as a score of data fit.25 This reveals a deep connection between
25 The remaining terms are independent of data. the rkhs norm and the associated Gaussian process: posterior mean
functions arising from “more unusual” observations from the point of
view of the gp have higher complexity from the point of view of the
rkhs. Whereas the Gaussian process judges the data y to be complex
26 The “needles in a haystack” constructed earlier via the marginal likelihood, the rkhs judges the resulting posterior mean
would have extraordinarily large rkhs norm, 𝜇D to be complex via the rkhs norm – but these are simply two ways of
if they lie in a given rkhs at all.
interpreting the same scenario.26
The role of the rkhs norm in quantifying function complexity sug-
gests a natural space of objective functions to work with when seeking
worst-case results (10.7). We take the rkhs ball of radius 𝐵, the space of
rkhs ball of radius 𝐵, H𝐾 [𝐵 ] functions with complexity bounded by 𝐵 in the rkhs:

H𝐾 [𝐵] = 𝑓 | k𝑓 k H𝐾 ≤ 𝐵 . (10.11)

The radius 𝐵 is left as a parameter that is absorbed into derived bounds.

10.3 relevant properties of covariance functions

A sizable majority of the results discussed in the remainder of this chap-
ter concern optimization performance on one of the function spaces
discussed in the previous section: sample paths of a centered Gaussian
process with covariance function 𝐾 or functions in the corresponding
rkhs H𝐾 . Given the fundamental role the covariance function plays
in determining sample path behavior, it should not be surprising that
the nature of the covariance function also has profound influence on
optimization performance.

220
 10.3. relevant properties of covariance functions

𝜈 = 1.01 𝜈 = 3/2 𝜈=2

Figure 10.3: Sample paths from a centered gp with Matérn covariance with smoothness parameter 𝜈 ranging from just
over 1 (left) to 2 (right). The random seed is shared so that paths of the same color are comparable. All samples
are once differentiable, but some are smoother than others.

One might intuitively expect that optimizing smoother functions

should be easier than optimizing rougher functions, as a rougher function
gives the optimum more places to “hide;” see the figure in the margin.
This intuition turns out to be correct. The key insight is that rougher
functions require more information to describe than smoother ones. As
each observation we make is limited in how much information it can
reveal regarding the objective function, rougher objectives require more
observations to learn with a similar level of confidence, and to optimize
with a similar level of success. We can make this intuition precise through
the concept of information capacity, which bounds the rate at which we
can learn about the objective through noisy observations and serves as Sample paths from a smooth gp (above) and a
a fundamental measure of function complexity appearing in numerous rough one (below). The rough samples have
analyses. more degrees of freedom – and far more local
maxima – and we might conclude they are
harder to optimize as a result.
Smoothness of sample paths
The connection between sample path smoothness and the inherent diffi- Matérn and squared exponential covariance
culty of learning is best understood for the Matérn covariance family and functions: § 3.3, p. 51
the limiting case of the squared exponential covariance. Combined, these
covariance functions allow us to model functions with a continuum of 27 The given expression has unit length and out-
smoothness. To this end, there is a general form of the Matérn covariance put scale; if desired, we can introduce param-
eters for these following § 3.4.
function modeling functions of any finite smoothness, controlled by a
parameter 𝜈 > 0:27 smoothness parameter, 𝜈

21−𝜈 √
𝜈 √

𝐾m (𝑑; 𝜈) = 2𝜈𝑑 𝐾𝜈 2𝜈𝑑 , (10.12)
Γ(𝜈)
where 𝑑 = |𝑥 − 𝑥 0 | and 𝐾𝜈 is the modified Bessel function of the sec-
ond kind. Sample paths from a centered Gaussian process with this
28 This can be made precise through the coeffi-
covariance are d𝜈e − 1 times continuously differentiable, but the smooth- cient of Hölder continuity in the “final” deriva-
ness of sample paths is not as granular as a simple count of derivatives. tive of 𝐾m , which controls the smoothness of
Rather, the parameter 𝜈 allows us to fine-tune sample path smoothness the final derivative of sample paths. Except
when 𝜈 is an integer, the Matérn covariance
as desired.28Figure 10.3 illustrates sample paths generated from a Matérn
with parameter 𝜈 belongs to the Hölder space
covariance with a range of smoothness from 𝜈 = 1.01 to 𝜈 = 2. All of
these samples are exactly once differentiable, but we might say that the C 𝛼,𝛽 ; 𝛼 = b2𝜈 c; 𝛽 = 2𝜈 − b2𝜈 c,
𝜈 = 1.01 samples are “just barely” so, and that the 𝜈 = 2 samples are and we can expect the final derivative of sam-
“very nearly” twice differentiable. ple paths to be (𝜈 − b𝜈 c)-Hölder continuous.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 221

 theoretical analysis

Taking the limit 𝜈 → ∞ recovers the squared exponential covariance

𝐾se . This serves as the extreme end of the continuum, modeling functions
with infinitely many continuous derivatives. Together, the Matérn and
squared exponential covariances allow us to model functions with any
Matérn family smoothness 𝜈 ∈ (0, ∞]; we will call this collection the Matérn family.

Information capacity
We now require some way to relate the complexity of sample path
behavior to our ability to learn about an unknown function. Information
theory provides an answer through the concept of information capacity,
the maximum rate of information transfer through a noisy observation
mechanism.
In the analysis of Bayesian optimization algorithms, the central con-
cern is how efficiently we can learn about a gp-distributed objective
function GP (𝑓 ; 𝜇, 𝐾) through a set of 𝜏 noisy observations D𝜏 . The infor-
mation capacity of this observation process, as a function of the number
of observations 𝜏, provides a fundamental bound on our ability to learn
about 𝑓. For this discussion, let us adopt the common observation model
of independent and homoskedastic additive Gaussian noise with scale
𝜎𝑛 > 0 (2.16). In this case, information capacity is a function of:
• the covariance 𝐾, which determines the information content of 𝑓, and
• the noise scale 𝜎𝑛 , which limits the amount of information obtainable
through a single observation.
29 We have 𝐻 [y | 𝑓 ] = 𝐻 [y | 𝝓 ] by conditional The information regarding 𝑓 contained in an arbitrary set of obser-
independence (1.3). vations D = (x, y) can be quantified by the mutual information (a.16):29
1
𝐼 (y; 𝑓 ) = 𝐻 [y] − 𝐻 [y | 𝝓] = 2 log |I + 𝜎𝑛−2 𝚺|, (10.13)

30 For more on information gain, see § 6.3, p. 115. where 𝚺 = 𝐾 (x, x). Note that the entropy of y given 𝝓 does not depend
Thus far we have primarily concerned our- on the actual value of 𝝓, and thus the mutual information 𝐼 (y; 𝑓 ) is also
selves with information gain regarding 𝑥 ∗ or
𝑓 ∗ ; here we are reasoning about the function
the information gain about 𝑓 provided by the data.30 The information ca-
𝑓 itself. pacity (also known as the maximum information gain) of this observation
process is now the maximum amount of information about 𝑓 obtainable
information capacity, 𝛾𝜏 through any set of 𝜏 observations:

𝛾𝜏 = sup 𝐼 (y; 𝑓 ). (10.14)

|x |=𝜏

Known bounds on information capacity

The information capacity of a gp observation process (10.14) is commonly
invoked in theoretical analyses of algorithms making use of this model
class. Unfortunately, working with information capacity is somewhat
unwieldy for two reasons. First, as mentioned above, information ca-
pacity is a function of the covariance function 𝐾, which can be verified
through the explicit formula in (10.13). Thus performance guarantees in
terms of information capacity require further analysis to derive explicit

222
 10.3. relevant properties of covariance functions

results for a particular choice of model. Second, information capacity of 31 c.-w. ko et al. (1995). An Exact Algorithm for
any given model is in general np-hard to compute due to the difficulty Maximum Entropy Sampling. Operations Re-
search 43(4):684–691.
of the set function maximization in (10.14).31
For these reasons, the typical strategy is to derive agnostic conver-
gence results in terms of the information capacity of an arbitrary obser-
vation process, then seek to derive bounds on the information capacity 32 n. srinivas et al. (2010). Gaussian Process Op-
for notable covariance functions such as those in the Matérn family. A timization in the Bandit Setting: No Regret
and Experimental Design. icml 2010.
common proof strategy for bounding the information gain is to relate the
information capacity to the spectrum of the covariance function, with 33 d. janz et al. (2020). Bandit optimisation of
functions in the Matérn kernel rkhs. aistats
faster spectral decay yielding stronger bounds on information capacity. 2020.
The first explicit bounds on information capacity for the Matérn and 34 s. vakili et al. (2021b). On Information Gain
squared exponential covariances were provided by srinivas et al.,32 and and Regret Bounds in Gaussian Process Ban-
the bounds for the Matérn covariance have since been sharpened using dits. aistats 2021.
similar techniques.33, 34
For a compact domain X ⊂ ℝ𝑑 and fixed noise scale 𝜎𝑛 , we have
the following asymptotic bounds on the information capacity. For the
Matérn covariance function with smoothness 𝜈, we have:34

𝑑
𝛾𝜏 = O 𝜏 𝛼 (log 𝜏) 1−𝛼 , 𝛼 = ; (10.15)
2𝜈 + 𝑑
and for the squared exponential covariance, we have:32


𝛾𝜏 = O (log 𝜏)𝑑+1 . (10.16)
These results embody our stated goal of characterizing smoother sample
paths (as measured by 𝜈) as being inherently less complex than rougher
sample paths. The information capacity decreases steadily as 𝜈 → ∞,
eventually dropping to only logarithmic growth in 𝜏 for the squared
exponential covariance. The correct interpretation of this result is that
the smoother sample paths require less information to describe, and
thus the maximum amount of information one could learn is limited
compared to rougher sample paths.

Bounding the sum of predictive variances

A key result linking information capacity to an optimization policy is the
following.32 Suppose some optimization policy selected an arbitrary se-
quence of 𝜏 observation locations {𝑥𝑖 }, and let {𝜎𝑖 } be the corresponding predictive standard deviation of 𝜙𝑖 , 𝜎𝑖
predictive standard deviations at the time of selection. By applying the
chain rule for mutual information, we can rewrite the information gain
(10.13) from observing y in terms of the marginal predictive variances:
1 ∑︁
𝜏  𝜎2 
𝐼 (y; 𝑓 ) = log 1 + 𝑖2 . (10.17)
2 𝑖=1 𝜎𝑛

Now assume that the prior covariance function is bounded: 𝐾 (𝑥, 𝑥) ≤ 𝑀. bound on prior variance 𝐾 (𝑥, 𝑥), 𝑀
Noting that 𝑧 2 / log(1 + 𝑧 2 ) is increasing for 𝑧 > 0 and that 𝜎𝑖2 ≤ 𝑀, the posterior variance is nonincreasing: § 2.2,
following inequality holds for every observation: p. 22.

𝑀  𝜎𝑖2 
𝜎𝑖2 ≤ log 1 + .
log(1 + 𝜎𝑛−2 𝑀) 𝜎𝑛2
Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 223
 theoretical analysis

We can now bound the sum of the predictive variances in terms of the
35 If an explicit leading constant is desired, we
Í information capacity:35
have 𝑖 𝜎𝑖2 ≤ 𝑐𝛾𝜏 with
2𝑀
∑︁
𝜏
𝑐= . 𝜎𝑖2 = O(𝛾𝜏 ). (10.18)
log(1 + 𝜎𝑛−2 𝑀)
𝑖=1

This bound will repeatedly prove useful below.

10.4 bayesian regret with observation noise

We have now covered the background required to understand – or at
least to meaningfully interpret – most of the theoretical convergence
results appearing in the literature. In the remainder of the chapter we will
summarize some notable results built upon these ideas. The literature is
expansive, and navigation can be challenging. Broadly, we can categorize
these results according to certain dichotomies in their approach and
focus, listed below.
analysis: frequentist or Bayesian? • The first is whether the result is regarding the average-case (Bayesian)
or worst-case (frequentist) regret. In the former case, we assume the
objective function is a sample path from a Gaussian process GP (𝑓 ; 𝜇, 𝐾)
and seek to bound the expected regret (10.6). In the latter case, we assume
the objective function lies in some rkhs H𝐾 with bounded norm (10.11)
and seek to bound the worst-case regret on this space (10.7).
observations: noisy or exact? • The second is the assumption and treatment of observation noise. Stronger
guarantees can often be derived in the noiseless setting, as we can learn
much faster about the objective function. Observation noise is also mod-
eled somewhat differently in the Bayesian and frequentist settings.
In this and the following sections, we will provide an overview of
convergence results for all combinations of these choices: frequentist
and Bayesian guarantees, with and without noise. For each of these cases,
we will discuss both upper bounds on regret, which provide guarantees
for the performance of specific Bayesian optimization algorithms, and
also lower bounds on regret, which provide algorithm-agnostic bounds
on the best possible performance. Here we will begin with results for
Bayesian regret in the noisy setting.
To facilitate the discussion, we will adopt the notation O∗ (some-
asymptotic behavior with logarithmic factors times written Õ in other texts) to describe asymptotic bounds in which
suppressed, O∗ dimension-independent logarithmic factors of 𝜏 are suppressed:



𝑓 (𝜏) = O 𝑔(𝜏) (log 𝜏)𝑘 ; =⇒ 𝑓 (𝜏) = O∗ 𝑔(𝜏) .

Common assumptions
assumption: 𝑓 is a sample path from In this section, we will assume that the objective function 𝑓 : X → ℝ is
GP (𝑓 ; 𝜇 ≡ 0, 𝐾) a sample path from a centered Gaussian process GP (𝑓 ; 𝜇 ≡ 0, 𝐾), and
assumption: observations are corrupted by iid that observation noise is independent, homoskedastic Gaussian noise
Gaussian noise with scale 𝜎𝑛 with scale 𝜎𝑛 > 0 (2.16). The domain X will at various times be either a

224
 10.4. bayesian regret with observation noise

finite set (as a stepping stone toward the continuous case) or a compact
and convex subset of a 𝑑-dimensional cube: X ⊂ [0, 𝑚]𝑑. We will also
be assuming that the covariance function is continuous and bounded
on X : 𝐾 (𝑥, 𝑥) ≤ 1. Since the covariance function is guaranteed to be assumption: 𝐾 (𝑥, 𝑥) ≤ 1 is bounded
bounded anyway (as X is compact) this simply fixes the scale without
loss of generality.

Upper confidence bound

In a landmark paper, srinivas et al. derived sublinear cumulative regret upper confidence bound: § 7.8, p. 145, § 8.4,
bounds for the Gaussian process upper confidence bound (gp-ucb) policy p. 170
in the Bayesian setting with noise.36 The authors considered policies of
the form (8.25): 36 n. srinivas et al. (2010). Gaussian Process Op-
timization in the Bandit Setting: No Regret
𝑥𝑖 = arg max 𝜇 + 𝛽𝑖 𝜎, (10.19)
and Experimental Design. icml 2010.
𝑥 ∈X

were 𝑥𝑖 is the point chosen in the 𝑖th iteration of the policy, 𝜇 and 𝜎 are
shorthand for the posterior mean and standard deviation of 𝜙 = 𝑓 (𝑥)
given the data available at time 𝑖, D𝑖−1 , and 𝛽𝑖 is a time-dependent
exploration parameter. The authors were able to demonstrate that if this
exploration parameter is carefully tuned over the course of optimization,
then the cumulative regret of the policy can be asymptotically bounded,
with high probability, in terms of the information capacity. information capacity: § 10.3, p. 222
We will discuss this result and its derivation in some detail below, as it
demonstrates important proof strategies that will be repeated throughout
this section and the next.

Regret bound on finite domains

To proceed, we first assume the domain X is finite; we will lift this to 37 The general strategy for this case was estab-
the continuous case shortly via a secondary argument.37 lished in the linear bandit setting in:
The construction of the ucb policy suggests the following confidence v. dani et al. (2008). Stochastic Linear Opti-
interval condition is likely to hold for any point 𝑥 ∈ X at any time 𝑖: mization Under Bandit Feedback. colt 2008.

𝜙 ∈ [𝜇 − 𝛽𝑖 𝜎, 𝜇 + 𝛽𝑖 𝜎] = B𝑖 (𝑥). (10.20)

In fact, we can show that for appropriately chosen confidence parameters step 1: show confidence intervals (10.20) are
{𝛽𝑖 }, every such confidence interval is always valid, with high probability. universally valid with high probability
This may seem like a strong claim, but it is simply a consequence of 38 srinivas et al. use
the exponentially decreasing tails of the Gaussian distribution. At time 𝑖,

we can use tail bounds on the Gaussian cdf to bound the probability of Pr 𝜙 ∉ B𝑖 (𝑥) ≤ exp(−𝛽𝑖2 /2).

a given confidence interval failing in terms of 𝛽𝑖 ,38 then use the union 39 Using the above bound, the probability of fail-
bound to bound the probability of (10.20) failing anywhere at time 𝑖.39 ure anywhere at time 𝑖 is at most
Finally, we show that by increasing the confidence parameter 𝛽𝑖 over
|X | exp(−𝛽𝑖2 /2).
time – so that the probability of failure decreases suitably quickly – the
probability of failure anywhere and at any time is small. srinivas et al. 40 The mysterious appearance of 𝜋 2 /6 comes
showed in particular that for any 𝛿 ∈ (0, 1), taking40 from
∑︁
∞
1 𝜋2
 𝑖 2 𝜋 2 |X |  𝑖 2
=
6
.
𝛽𝑖2 = 2 log (10.21) 𝑖=1
6𝛿
Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 225
 theoretical analysis

guarantees that the confidence intervals (10.20) are universally valid

with probability at least 1 − 𝛿.
step 2: bound instantaneous regret by width With this, we are actually almost finished. The next key insight is
of chosen confidence interval that if the confidence intervals in (10.20) are universally valid, then we
can bound the instantaneous regret of the policy in every iteration by
noting that the confidence interval of the chosen point always contains
the global optimum:
𝑓 ∗ ∈ B𝑖 (𝑥𝑖 ). (10.22)
the lower bound of every point holds for 𝑓 ∗ To show this, we first note that the lower bound of every confidence
interval applies to 𝑓 ,∗ as all intervals are valid and 𝑓 ∗ is the global maxi-
the upper bound of the chosen point holds for mum. Further, the upper bound of the chosen confidence interval is valid
every function value for 𝑓 ,∗ as it is the maximal upper bound by definition (and thus is valid
for every function value) (10.19).
This result allows us to bound, with high probability, the instan-
taneous regret (10.2) in every iteration by the width of the confidence
interval of the chosen point:

𝜌𝑖 ≤ 2𝛽𝑖 𝜎𝑖 . (10.23)

41 d. russo and b. van roy (2014). Learning to russo and van roy interpret this bound on the instantaneous regret as
Optimize via Posterior Sampling. Mathematics
of Operations Research 39(4):1221–1243.
guaranteeing that regret can only be high when we also learn a great
deal about the objective function to compensate (10.17).41
step 3: bound cumulative regret via bound on Finally, we bound the cumulative regret. Assuming the confidence
instantaneous regret intervals (10.20) are universally valid, we may bound the sum of the
squared instantaneous regret up to time 𝜏 by

∑︁
𝜏 ∑︁
𝜏 ∑︁
𝜏
𝜌𝑖2 ≤ 4 𝛽𝑖2𝜎𝑖2 ≤ 4𝛽𝜏2 𝜎𝑖2 = O(𝛽𝜏2 𝛾𝜏 ). (10.24)
𝑖=1 𝑖=1 𝑖=1

From left-to-right, we plug in (10.23), note that {𝛽𝑖 } is nondecreasing

(10.21), and appeal to the information capacity bound on the sum of
42 We have 𝑀 = 1 as a bound on 𝐾 (𝑥, 𝑥) accord- predictive variances (10.18).42 Plugging in 𝛽𝜏 and appealing to the Cauchy–
ing to our common assumptions (p. 224).
Schwartz inequality gives
√︁
𝑅𝜏 = O∗ ( 𝜏𝛾𝜏 log |X |) (10.25)

with probability at least 1 − 𝛿.

Extending to continuous domains

The above analysis can be extended to continuous domains via a dis-
cretization argument. The proof is technical, but the technique is often
useful in other settings, so we provide a sketch of a general argument:
assumption: X is convex and compact • We assume that the domain X ⊂ [0, 𝑚]𝑑 is convex and compact.
assumption: 𝑓 is Lipschitz continuous • We assume that the objective function is continuously differentiable. As
X is compact, this implies the objective is in fact Lipschitz continuous
with some Lipschitz constant 𝐿.

226
 10.4. bayesian regret with observation noise

• Purely for the sake of analysis, in each iteration 𝑖, we discretize the step 1: discretize the domain with a sequence
domain with a grid X𝑖 ⊂ X ; these grids become finer over time and of increasingly fine grids {𝑋𝑖 }
eventually dense. The exact details vary, but it is typical to take a regular
grid in [0, 𝑚]𝑑 with spacing on the order of O(1/𝑖 𝑐 ) (for some constant
𝑐 not depending on 𝑑) and take the intersection with X. The resulting
discretizations have size log |X𝑖 | = O(𝑑 log 𝑖). size of discretizations: log |X𝑖 | = O (𝑑 log 𝑖)
• We note that Lipschitz continuity of 𝑓 allows us to extend valid confi- step 2: valid confidence intervals on X𝑖 can
dence intervals for the function values at X𝑖 to all of X with only slight be extended to all of X with slight inflation
inflation. Namely, for any 𝑥 ∈ X, let [𝑥] 𝑖 denote the closest point to 𝑥
in X𝑖 . By Lipschitz continuity, a valid confidence interval at [𝑥] 𝑖 can be
extended to one at 𝑥 with inflation on the order of O(𝐿/𝑖 𝑐 ) due to the inflation in iteration 𝑖: O (𝐿/𝑖 𝑐 )
discretizations becoming finer over time.
• With this intuition, we design a confidence parameter sequence {𝛽𝑖 } step 3: find a confidence parameter sequence
guaranteeing that the function values on both the grids {X𝑖 } and at {𝛽𝑖 } guaranteeing validity on {𝑥𝑖 } and {X𝑖 }
the points chosen by the algorithm {𝑥𝑖 } always lie in their respective
confidence intervals with high probability. As everything is discrete, we
can generally start from a guarantee such as (10.21), replace |X | with
|X𝑖 |, and fiddle with constants as necessary.
• Finally, we proceed as in the finite case. We bound the instantaneous step 4: proceed as in the discrete case
regret in each iteration in terms of the width of the confidence intervals
of the selected points, noting that any extra regret due to discretization
shrinks rapidly as O(𝐿/𝑖 𝑐 ) and, if we are careful, does not affect the
asymptotic regret. Generally the resulting bound simply replaces any
factors of log |X | with factors of log |X𝑖 | = O(𝑑 log 𝑖).
For the particular case of bounding the Bayesian regret of gp-ucb,
we can effectively follow the above argument but must deal with some
nuance regarding the Lipschitz constant of the objective function. First,
note that if the covariance function of our Gaussian process is smooth sample path differentiability: § 2.6, p. 30
enough, its sample paths will be continuously differentiable43 and Lips-
chitz continuous – but with random Lipschitz constant. srinivas et al. 43 Hölder continuity of the derivative process co-
proceed by assuming that the covariance function is sufficiently smooth variance is sufficient; this holds for the Matérn
family with 𝜈 > 1.
to ensure that the Lipschitz constant has an exponential tail bound of
the following form:
∀𝜆 > 0 : Pr(𝐿 > 𝜆) ≤ 𝑑𝑎 exp(−𝜆 2 /𝑏 2 ), (10.26)
where 𝑎, 𝑏 > 0; this allows us to derive a high-probability upper bound
on 𝐿. We will say more about this assumption shortly.
srinivas et al. showed that, under this assumption, a slight modifica-
tion of the confidence parameter sequence from the finite case44 (10.21) 44 Specifically, srinivas et al. took:
is sufficient to ensure √︁  2𝑖 2 𝜋 2 
𝑅𝜏 = O∗ ( 𝑑𝜏𝛾𝜏 ) (10.27) 𝛽𝑖2 = 2 log
3𝛿
√︁

with high probability. Thus, assuming the Gaussian process sample paths + 2𝑑 log 𝑖 2𝑑𝑏𝑚 log(4𝑑𝑎/𝛿) ;
are smooth enough, the cumulative regret of gp-ucb on continuous The resulting bound holds with probability at
domains grows at rate comparable to the discrete case. least 1 − 𝛿.
Plugging in
√ the information capacity bounds in (10.15–10.16) and
dropping the 𝑑 factor, we have the following high-probability regret

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 227

 theoretical analysis

bounds for specific covariance functions. For the Matérn covariance, we

have

𝜈 +𝑑 2𝜈
𝑅𝜏 = O∗ 𝜏 𝛼 (log 𝜏) 𝛽 , 𝛼= , 𝛽= , (10.28)
2𝜈 + 𝑑 2𝜈 + 𝑑
and for the squared exponential we have
√︁

𝑅𝜏 = O∗ 𝜏 (log 𝜏)𝑑 . (10.29)

The growth is sublinear for all values of the smoothness parameter 𝜈, so

gp-ucb achieves no regret with high probability for the Matérn family.

The Lipschitz tail bound condition

Finally, we address the exponential tail bound assumption on the Lip-
schitz constant (10.26). Despite its seeming strength in controlling the
behavior of sample paths, it is actually a fairly weak assumption on
the Gaussian process. In fact, all that is needed is that sample paths be
continuously differentiable. In this case, each coordinate of the gradient,
45 a. w. van der vaart and j. a. wellner (1996). 𝑓𝑖 = 𝜕𝑓 /𝜕𝑥𝑖 , is a sample path continuous Gaussian process. By compact-
Weak Convergence and Empirical Processes ness, each 𝑓𝑖 is then almost surely bounded on X. Now the Borell–tis
with Applications to Statistics. Springer–Verlag.
[proposition a.2.1]
inequality ensures that for any 𝜆 > 0 we have:45

√  
Pr max |𝑓𝑖 | > 𝜆 ≤ 4 exp(−𝜆 2 /𝑏𝑖2 ), 𝑏𝑖 = 2 2 𝔼 max |𝑓𝑖 | . (10.30)

Taking a union bound then establishes a bound of the desired form

(10.26). In particular, this argument applies to the Matérn covariance
with 𝜈 > 1, which has continuously differentiable sample paths. Thus
gp-ucb can achieve sublinear cumulative regret for all but the roughest
of sample paths.
However, the bound in (10.30) does not immediately lead to an algo-
rithm we can realize in practice, as the expected extremum 𝔼[max |𝑓𝑖 |]
is difficult to compute (or even bound). Things simplify considerably
if sample paths are twice differentiable, in which case ghosal and roy
46 s. ghosal and a. roy (2006). Posterior Con- showed how to derive bounds of the above form (10.30) for each of the co-
sistency of Gaussian Process Prior for Non- ordinates of the gradient process with explicitly computable constants,46
parametric Binary Regression. The Annals of
Statistics 34(5):2413–2429. [lemma 5]
yielding a practical algorithm for the Matérn covariance with 𝜈 > 2.
Fortunately, intricate arguments regarding the objective function’s
47 n. de freitas et al. (2012a). Regret Bounds Lipschitz constant are only necessary due to its randomness in the Bayes-
for Deterministic Gaussian Process Bandits. ian setting. In the frequentist setting, where 𝑓 is in some rkhs ball H𝐾 [𝐵],
arXiv: 1203.2177 [cs.LG] [lemma 1]
we can immediately derive a hard upper bound 𝐿 in terms of 𝐾 and 𝐵.47

Bounding the expected regret

Without too much effort, we may augment the high-probability bound
presented above with a corresponding bound on the expected regret
(10.6). However, note that the high-probability regret bound is stronger in
the sense that we can guarantee good performance not only on average
but also in extremely “unlucky” scenarios as well.

228
 10.4. bayesian regret with observation noise

For simplicity, let us begin again with a finite domain X. Consider

a run of the gp-ucb algorithm in (10.19) for some confidence parameter
sequence {𝛽𝑖 } to be determined shortly. Let {𝑥𝑖 } be the sequence of
selected points and let 𝑥 ∗ be a global optimum of 𝑓. Define

𝑢𝑖 = 𝜇𝑖 + 𝛽𝑖 𝜎𝑖 ; 𝑢𝑖∗ = 𝜇𝑖∗ + 𝛽𝑖 𝜎𝑖∗

to be, respectively, the upper confidence bound associated with 𝑥𝑖 at the

time of its selection and the upper confidence bound associated with
𝑥 ∗ at the same time. Since the algorithm always maximizes the upper
confidence bound in each iteration, we always have 𝑢𝑖 ≤ 𝑢𝑖∗ . This allows
us to bound the expected regret as follows:
∑︁
𝜏 ∑︁
𝜏 ∑︁
𝜏
𝔼[𝑅𝜏 ] = 𝔼[𝑓 ∗ − 𝜙𝑖 ] ≤ 𝔼[𝑓 ∗ − 𝑢𝑖∗ ] + 𝔼[𝑢𝑖 − 𝜙𝑖 ], (10.31)
𝑖=1 𝑖=1 𝑖=1

where we have added 𝑢𝑖 − 𝑢𝑖∗ ≥ 0 to each term and rearranged.

In their analysis of the Gaussian process Thompson sampling al-
gorithm (discussed below), russo and van roy showed how to bound
each of the terms on the right-hand side. First, they show that the confi-
dence parameter sequence (compare with the analogous and spiritually 48 This is again a consequence of the rapidly de-
equivalent sequence used above (10.21)) caying tails of the Gaussian distribution; note
that 𝑓 ∗ − 𝑢𝑖∗ has distribution
 (𝑖 2 + 1)|X | 

N −𝛽𝑖 𝜎𝑖∗ , (𝜎𝑖∗ ) 2 .
𝛽𝑖2 = 2 log √
2𝜋 As 𝛽𝑖 increases without bound, this term will
Í eventually be negative with overwhelming
is sufficient to bound the first term by a constant: ≤ 1.48 The ∗ ∗
𝑖 𝔼[𝑓 −𝑢𝑖 ] probability.
second term can then be bounded in terms of the information capacity
following our previous discussion (10.18, 10.24):49 49 𝑢𝑖 − 𝜙𝑖 has distribution N (𝛽𝑖 𝜎𝑖 , 𝜎𝑖2 ).

∑︁
𝜏 ∑︁
𝜏 ∑︁
𝜏 √︁
𝔼[𝑢𝑖 − 𝜙𝑖 ] = 𝛽𝑖 𝜎𝑖 ≤ 𝛽𝜏 𝜎𝑖 = O∗ ( 𝜏𝛾𝜏 log|X |). (10.32) 50 Although used for Thompson sampling, the
discretization argument used in the below ref-
𝑖=1 𝑖=1 𝑖=1
erence suffices here as well:
As before, we may extend this result to the √︁
continuous case via a k. kandasamy et al. (2018). Parallelised Bayes-
discretization argument to prove 𝔼[𝑅𝜏 ] = O∗ ( 𝑑𝜏𝛾𝜏 ), matching the ian Optimisation via Thompson Sampling. ais-
high-probability bound (10.27).50 tats 2018 [appendix, theorem 11].

Thompson sampling
russo and van roy developed a general approach for transforming Thompson sampling: § 7.9, p. 148, § 8.7, p. 176
Bayesian regret bounds for a wide class of ucb-style algorithms into
regret bounds for analogous Thompson sampling algorithms.51 51 d. russo and b. van roy (2014). Learning to
Namely, consider a ucb policy selecting a sequence of points {𝑥𝑖 } Optimize via Posterior Sampling. Mathematics
of Operations Research 39(4):1221–1243.
for observation by maximizing a sequence of “upper confidence bounds,”
which here can be any deterministic functions of the observed data,
regardless of their statistical validity. Let {𝑢𝑖 } be the sequence of upper
confidence bounds associated with the selected points at the time of
their selection, and let {𝑢𝑖∗ } be the sequence of upper confidence bounds
associated with a given global optimum, 𝑥 ∗.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 229

 theoretical analysis

Now consider a corresponding Thompson sampling policy selecting

𝑥𝑖 ∼ 𝑝 (𝑥 ∗ | D𝑖−1 ). Note that, given the observed data, 𝑥𝑖 and 𝑥 ∗ are
identically distributed by design; because the upper confidence bounds
are deterministic functions of the observed data, 𝑢𝑖 and 𝑢𝑖∗ are thus
identically distributed as well. This allows us to express the expected
cumulative regret of the Thompson sampling policy entirely in terms of
the upper confidence bounds:
∑︁
𝜏 ∑︁
𝜏
𝔼[𝑅𝜏 ] = 𝔼[𝑓 ∗ − 𝑢𝑖∗ ] + 𝔼[𝑢𝑖 − 𝜙𝑖 ]; (10.33)
𝑖=1 𝑖=1

This is of the exactly same form as the bound derived above for gp–
ucb (10.31), leading immediately to a regret bound for gp-ts matching
that for gp-ucb (10.32):
√︁
𝔼[𝑅𝜏 ] = O∗ ( 𝜏𝛾𝜏 log|X |); (10.34)

52 k. kandasamy et al. (2018). Parallelised Bayes- once again, we may extend this result to the continuous case to derive a
ian Optimisation via Thompson Sampling. ais- bound matching that for gp-ucb (10.27).52
tats 2018.

Upper bounds on simple regret

𝑟𝜏 ≤ 𝑅𝜏
𝜏 We may use the bound on simple regret in terms of the average regret in
(10.4) to derive bounds on the simple regret of gp-ucb and gp-ts policies.
53 There
√︁ is an additional multiplicative
√ factor of Dropping dependence on the domain size,53 we have
log |X | in the finite case and 𝑑 in the con- √︁

tinuous case. 𝑟𝜏 = O∗ 𝛾𝜏 /𝜏

for both algorithms, where this is to be understood as either a high-

probability (gp-ucb) or expected-case (gp-ucb, gp-ts) result.
Plugging in the information capacity bounds (10.15–10.16), we have
the following bounds for the simple regret for specific covariance func-
tions. For the Matérn covariance in dimension 𝑑, we have

𝜈 2𝜈
𝑟𝜏 = O∗ 𝜏 𝛼 (log 𝜏) 𝛽 , 𝛼 =− , 𝛽= ,
2𝜈 + 𝑑 2𝜈 + 𝑑
and for the squared exponential we have
√︁

𝑟𝜏 = O∗ (log 𝜏)𝑑 /𝜏 .

Lower bounds and tightness of existing algorithms

We have now derived upper bounds on the regret of particular Bayesian
optimization algorithms in the Bayesian setting with noise. A natural
question is whether we can derive corresponding algorithm-agnostic
lower regret bounds establishing the fundamental difficulty of optimiza-
tion in this setting, which might hint at how much room for improvement
there may be.
Lower bounds on Bayesian regret are not easy to come by. As we
will see, the frequentist setting offers considerable flexibility in deriving

230
 10.4. bayesian regret with observation noise

𝑓− 𝑓 𝑓+
Figure 10.4: A sketch of scarlett’s
proof strategy. Given access
to a reference function 𝑓,
which of its translations by
Δ, 𝑓 − or 𝑓,+ is the objective
function?

Δ Δ

lower bounds, as we can construct explicit objective functions in a given

rkhs, then prove that they are difficult to optimize. In the Bayesian
setting, the objective function is random, so we must instead seek explicit
distributions over objective functions with enough structure that we can
bound the expected regret, a much more challenging task.
That said, nontrivial lower bounds have been derived in this setting,
most notably on the unit interval X = [0, 1]. Under the assumption of
a stationary Gaussian process with twice differentiable sample paths, 54 j. scarlett (2018). Tight Regret Bounds for
scarlett demonstrated the following high-probability lower bound on Bayesian Optimization in One Dimension.
icml 2018.
the expected cumulative regret of any optimization algorithm:54
√
𝔼[𝑅𝜏 ] = Ω( 𝜏). (10.35)
This result is as about as good as we could hope for, as it matches the 55 o. shamir (2013). On the Complexity of Bandit
lower bound for optimizing convex functions.55 and Derivative-Free Stochastic Convex Opti-
mization. colt 2013.
The key idea behind this result is to identify multiple plausible ob-
jective functions with identical distributions such that an optimization
algorithm is prone to “prefer the wrong function,” and in doing so, incur
high regret. To illustrate the construction, we imagine an objective func-
tion on the unit interval is generated indirectly by the following process.
We first realize an initial sample path 𝑓 on the larger domain [−Δ, 1 + Δ],
56 This introduces the additional minor require-
for some Δ > 0.56 We then take the objective function to be one of the ment that the covariance function be defined
following translations of 𝑓 with equal probability: on this larger domain.

𝑓 + : 𝑥 ↦→ 𝑓 (𝑥 + Δ); 𝑓 − : 𝑥 ↦→ 𝑓 (𝑥 − Δ). (10.36) 57 This is a so-called genie argument: as the extra

information (provided by a “genie”) can be ig-
As the prior process is assumed to be stationary, this procedure does not nored, it cannot possibly impede optimization.
change the distribution of the objective function. 58 j. scarlett and v. cevhar (2021). An Introduc-
To proceed, we consider optimization of the generated objective given tory Guide to Fano’s Inequality with Applica-
tions in Statistical Estimation. In: Information-
access to the initial reference function 𝑓.57 We now frame optimization in Theoretic Methods in Data Science.
terms of using a sequence of noisy observations to determine which of 59 Specifically, we need that the global maximum
the two possible translations (10.36) represents the objective function; of the reference function is unlikely to be
see figure 10.4. Fano’s inequality,58 an information-theoretic lower bound “pushed off the edge” of the domain during
on error probability in adaptive hypothesis testing, then allows us to translation, and that sample paths are likely to
have locally quadratic behavior in a neighbor-
show that there is a significant probability that the data collected by any hood of the global optimum. The lower bound
algorithm have better cumulative regret for the wrong translation. Finally, on regret then holds with probability depend-
we may use assumed statistical properties of the objective function to ing on these events. For more discussion, see:
show that when this happens, the cumulative regret is significant.59 n. de freitas et al. (2012b). Exponential Re-
On the rougher side of the spectrum, wang et al. used the same proof gret Bounds for Gaussian Process Bandits with
strategy to bound the expected regret – both simple and cumulative – of Deterministic Observations. icml 2012.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 231

 theoretical analysis

60 z. wang et al. (2020b). Tight Regret Bounds any algorithm maximizing a (nondifferentiable) sample path of Brownian
for Noisy Optimization of a Brownian Motion. motion (from the Wiener process) on the unit interval:60
arXiv: 2001.09327 [cs.LG].
√︁ √︁
𝔼[𝑟𝜏 ] = Ω(1/ 𝜏 log 𝜏); 𝔼[𝑅𝜏 ] = Ω( 𝜏/log 𝜏). (10.37)

Both wang et al. and scarlett also described straightforward, but

not necessarily practical, optimization algorithms to provide correspond-
ing upper bounds on the unit interval. The algorithms are based on a
simple branch-and-bound scheme whereby the domain is adaptively
partitioned based on confidence bounds computed from available data.
61 Again 𝜈 > 2 is sufficient for the Matérn family. For relatively smooth sample paths,61 scarlett’s algorithm achieves
cumulative regret √︁
𝔼[𝑅𝜏 ] = O( 𝜏 log 𝜏)
√︁
with high probability. This is within a factor of log 𝜏 of the correspond-
ing lower bound (10.35), so there is not too much room for improvement
on the unit interval. Note that both gp-ucb and gp-ts with the squared
exponential covariance match this rate up to logarithmic factors (10.29),
but scarlett’s algorithm is better for the Matérn covariance for 𝜈 > 2
(10.28) (as assumed in the analysis). For Brownian motion sample paths,
wang et al. established the following upper bounds:
√ √
𝔼[𝑟𝜏 ] = O(log 𝜏/ 𝜏); 𝔼[𝑅𝜏 ] = O( 𝜏 log 𝜏), (10.38)

which are within a factor of (log 𝜏) 3/2 of the corresponding lower bounds
(10.37), again fairly tight.

10.5 worst-case regret with observation noise

We now turn our attention to worst-case (frequentist) results analogous
to the expected-case (Bayesian) results discussed in the previous section.
Here we no longer reason about the objective function as a sample
path from a Gaussian process, but rather as an fixed, but unknown
function lying in some reproducing kernel Hilbert space H𝐾 with norm
bounded by some constant 𝐵 (10.11). We also replace the assumption of
independent Gaussian observation errors with somewhat more flexible
and agnostic conditions. The goal is then to bound the worst-case (simple
or cumulative) regret (10.7) of a given algorithm when applied to such a
worst-case regret on H𝐾 [𝐵 ]: 𝑟¯𝜏 [𝐵 ], 𝑅¯𝜏 [𝐵 ] function, which we will notate with 𝑟¯𝜏 [𝐵] and 𝑅¯𝜏 [𝐵], respectively.
Compared with the analysis of Bayesian regret, the analysis of worst-
62 Recall that sample paths of a centered Gaus- case regret is complicated by model misspecification, as the model
sian process with covariance 𝐾 do not lie in- assumptions underlying the Bayesian optimization algorithms are no
side H𝐾 ; see p. 219.
longer valid.62 We must therefore seek other methods of analysis.

Common assumptions
assumption: 𝑓 ∈ H𝐾 [𝐵 ] In this section, we will assume that the objective function 𝑓 : X → ℝ,
assumption: X is compact where X is compact, lies in a reproducing kernel Hilbert space corre-
sponding to covariance function 𝐾 and has bounded norm: 𝑓 ∈ H𝐾 [𝐵]

232
 10.5. worst-case regret with observation noise

(10.11). As in the previous section, we will also assume that the covariance assumption: 𝐾 (𝑥, 𝑥) ≤ 1 is bounded
function is continuous and bounded on X : 𝐾 (𝑥, 𝑥) ≤ 1.
To model observations, we will assume that a sequence of observa- assumption: errors have zero mean
tions {𝑦𝑖 } at {𝑥𝑖 } are corrupted by additive noise, 𝑦𝑖 = 𝜙𝑖 + 𝜀𝑖 , and that conditioned on their history
the distribution of the errors satisfies mild regularity conditions. First,
we will assume each 𝜀𝑖 has mean zero conditioned on its history:
𝔼[𝜀𝑖 | 𝜺 <𝑖 ] = 0,
where 𝜺 <𝑖 is the vector of errors occurring before time 𝑖. We will also assumption: scale of errors is limited
make assumptions regarding the scale of the errors. The most typical
assumption is that the distribution of each 𝜀𝑖 is 𝜎𝑛 -sub-Gaussian condi- sub-Gaussian distribution
tioned on its history, that is, that it the tail of the conditional distribution
shrinks at least as quickly as a Gaussian distribution with variance 𝜎𝑛2 :



∀𝑐 > 0 : Pr |𝜀𝑖 | > 𝑐 | 𝜺 <𝑖 ≤ 2 exp − 12 𝑐 2 /𝜎𝑛2 .
This condition is satisfied, for example, by a distribution bounded on
the interval [−𝜎𝑛 , 𝜎𝑛 ] and by any Gaussian distribution with standard
deviation of at most 𝜎𝑛 .
Complementary with the above assumptions, the Bayesian optimiza-
tion algorithms that we will analyze model the function with the centered
Gaussian process GP (𝑓 ; 𝜇 ≡ 0, 𝐾) and assume independent Gaussian
observation noise with scale 𝜎𝑛 .

Upper confidence bound and Thompson sampling

Sublinear bounds on cumulative regret have been established for both
the Gaussian process upper confidence bound (gp-ucb) and Thompson
sampling (gp-ts) algorithms in this setting, albeit with somewhat slower
convergence rates that in the Bayesian setting. However, more complex
algorithms built on similar ideas are able to close this gap, as we will
discuss shortly.
The primary proof strategy in this setting to bound the deviation
of the Gaussian process posterior mean used in the algorithm from the
objective function in the assumed rkhs. A prototypical result is to show
that for some sequence of confidence parameters {𝛽𝑖 }, the confidence
interval assumption (10.20) is universally valid with high probability:
𝜙 ∈ [𝜇 − 𝛽𝑖 𝜎, 𝜇 + 𝛽𝑖 𝜎]. (10.39)
This would then allow us to bound the cumulative regret in terms of the finite-domain Bayesian analysis of gp-ucb:
information capacity following our previous analysis. p. 225
The strongest known result of this form was derived by abbasi-
yadkori in the context of regression63 and later applied in the context of
63 y. abbasi-yadkori (2012). Online Learning
optimization by chowdhury and gopalan.64 Namely, under our com- for Linearly Parameterized Control Problems.
mon assumptions for this section and the assumption of 𝜎𝑛 -sub-Gaussian PhD thesis. University of Alberta. [theorem
errors, the following confidence parameter sequence yields universally 3.11, remark 3.13]
valid confidence intervals with probability at least 1 − 𝛿: 64 s. r. chowdhury and a. gopalan (2017). On
√ Kernelized Multi-armed Bandits. icml 2017.
𝛽𝑖 = 𝐵 + 𝜎𝑛 2𝛾𝑖−1 + 2 log(1/𝛿). (10.40) [theorem 2]

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 233

 theoretical analysis

Unfortunately, these confidence parameters are much larger than

needed in the Bayesian setting (see (10.21) and footnote 44), and the
65 If we wish to retain explicit dependence on resulting regret bounds suffer as a result. Following the same argument
the rkhs norm 𝐵, we have leading up to (10.24), we may show that, with high probability, the worst-
√ √

𝑅¯𝜏 [𝐵 ] = O∗ 𝜏𝛾𝜏 (𝐵 + 𝛾𝜏 ) . case cumulative regret of gp-ucb is65
√
𝑅¯𝜏 [𝐵] = O∗ ( 𝜏𝛾𝜏 ); (10.41)
√
extending to continuous domains with a the same bound holds for gp-ts with an additional factor of 𝑑 stemming
discretization argument: p. 226 from a discretization argument.66
√
There is a gap on the order of 𝛾𝜏 between these bounds and the
66 s. r. chowdhury and a. gopalan (2017). On analogous bounds in the Bayesian setting (10.27), and for the Matérn
Kernelized Multi-armed Bandits. icml 2017. family, the best known bounds on the information capacity (10.15–10.16)
only guarantee sublinear cumulative regret when 𝑑 < 2𝜈. It is unclear
whether this gap stems from inherent weakness of the gp-ucb and gp-ts
algorithms in this setting or merely from weakness in their analysis to
67 s. vakili et al. (2021c). Open Problem: Tight date, and there may be room for improvement by proving that a tighter
Online Confidence Intervals for rkhs Ele- confidence parameter sequence than above (10.40) would suffice for
ments. colt 2021.
universally valid confidence intervals.67

Simple regret bounds in the nonadaptive setting

Although the best (yet) known worst-case regret bounds for the vanilla
gp-ucb and gp-ts algorithms do not grow favorably compared with their
Bayesian regret counterparts, several authors have been able to construct
spiritually related algorithms that close this gap.
A common idea in the development of these algorithms is to carefully
sever some of the dependence between the sequence of points observed
throughout optimization {𝑥𝑖 } and the corresponding observed values
{𝑦𝑖 }. This may seem counterintuitive, as the power of Bayesian optimiza-
tion algorithms surely stems from their adaptivity! However, injecting
some independence allows us to appeal to significantly more powerful
concentration results in their analysis.
The following result illustrates the power of nonadaptivity in this
68 A wider family of light-tailed noise distribu-
setting. Consider conditioning on a an arbitrary sequence of observation
tions was also considered.
locations {𝑥𝑖 } chosen independently of the corresponding values {𝑦𝑖 }.
69 s. vakili et al. (2021a). Optimal Order Simple
Regret for Gaussian Process Bandits. neurips Under our common assumptions for this section and the assumption of
2021. [theorem 1] 𝜎𝑛 -sub-Gaussian errors,68 vakili et al. showed that for any 𝑥 ∈ X, the
70 If X is finite, we can ensure universally valid following confidence interval condition holds with probability at least
confidence intervals at all times with proba- 1 − 𝛿:69
bility 1 − 𝛿 as in (10.21): √︁
√︃ 𝜙 ∈ [𝜇 − 𝛽 𝜎, 𝜇 + 𝛽 𝜎]; 𝛽 = 𝐵 + 𝜎𝑛 2 log(1/𝛿). (10.42)
𝛽𝑖 = 𝐵 + 𝜎𝑛 2 log(𝑖 2 𝜋 2 |X |/6𝛿).
Although this bound only holds for a fixed point at a fixed time, we may
For continuous domains, we can again rely extend this result to hold at all points and at all times with only logarith-
on discretization arguments. For a convex and
compact domain X ⊂ 𝑅𝑑 coupled with a co-
mic inflation of the confidence parameter.70 Comparing to chowdhury
variance function in the Matérn family with and gopalan’s corresponding result in the adaptive setting (10.40), the
𝜈 > 1, we may take: confidence parameter sequence here does not depend on the information
√︁ capacity. Thus we can ensure dramatically tighter confidence intervals
𝛽𝑖 = 𝐵 + 𝜎𝑛 𝑑 log(𝑖/𝛿).
when observations are designed nonadaptively.

234
 10.5. worst-case regret with observation noise

This result has enabled several strong convergence guarantees in

the worst-case setting with noise. vakili et al., for example, were able
to show that the simple policy of uncertainty sampling (also known as
maximum variance reduction) achieves near-optimal simple regret in the
worst case.71 Uncertainty sampling designs each observation by maximiz- 71 This result uses the alternative definition of
ing the posterior predictive variance, which for a Gaussian process does simple regret mentioned in footnote 5: the dif-
ference between the global maximum and the
not depend on the observed values (2.19). One insightful interpretation maximum of the posterior mean.
of uncertainty sampling is that, from the perspective of the Gaussian
process used in the algorithm, it acts to greedily maximize the informa-
tion about the objective function revealed by the observations (10.17) –
such a policy is sometimes described as performing “pure exploration.”
For typical (stationary) processes on typical (convex and compact)
domains, uncertainty sampling designs observations seeking to, in a
rough sense, cover the domain as evenly as possible; an example run
is shown in the marginal figure. In light of the concentration result
above, we might expect this space-filling behavior to guarantee that the
maximum of the posterior mean cannot stray too far from the maximum
of the objective function. vakili et al. showed that this is indeed the
case for the Matérn family. Namely, under our shared assumptions for
this section and the minor additional assumption that X is convex, the A sequence of 20 observations designed
simple regret of uncertainty sampling is, with high probability, bounded via a nonadaptive uncertainty sampling
by72 policy for the running example from chapter 7.
√︁
𝑟¯𝜏 [𝐵] = O∗ ( 𝑑𝛾𝜏 /𝜏).

This rate is within logarithmic factors of the best-known lower bounds 72 s. vakili et al. (2021a). Optimal Order Simple
for convergence in simple regret, as discussed later in this section. Regret for Gaussian Process Bandits. neurips
2021. [theorem 3, remark 2]

Closing the cumulative regret bound gap

Although uncertainty sampling achieves near-optimal simple regret in
the worst case, we cannot expect to meaningfully bound its cumulative
regret due to its nonadaptive nature. However, the results discussed
above have proven useful for building algorithms that do achieve near-
optimal cumulative regret in the worst case. To make the best use of
results such as vakili et al.’s towards this end, algorithm development 73 z. li and j. scarlett (2021). Gaussian Process
becomes a careful balancing act of including enough adaptivity to ensure Bandit Optimization with Few Batches. arXiv:
low cumulative regret while including enough independence to ensure 2110.07788 [stat.ML].
that we can prove low cumulative regret.
li and scarlett proposed and analyzed one relatively simple algo-
rithm in this direction.73 Given an observation budget of 𝜏, we proceed by
constructing a sequence of batch observations in a nonadaptive fashion.
After each batch observation is resolved, we pause to derive the resulting
confidence intervals from the observed data and eliminate any points
from consideration whose upper confidence bound is dominated by the The lighter blue region of this Gaussian
maximum lower confidence bound among the points remaining under process has upper confidence bounds dom-
inated by the maximum lower confidence
consideration; see the marginal figure. This occasional elimination step bound (dashed line). Assuming the confidence
is the only time the algorithm exhibits adaptive behavior – in fact, after bounds are valid with high probability, points
every elimination step, the gp used in the algorithm is “reset” to the prior in this region are unlikely to be optimal.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 235

 theoretical analysis

(on a smaller domain) by discarding all observed data. This ensures that
every observation is always gathered nonadaptively.
The intervening batches are constructed by uncertainty sampling.
The motivation behind this strategy is that we can bound the width of
74 x. cai et al. (2021). Lenient Regret and Good- confidence intervals after 𝑖 rounds of uncertainty sampling in terms of the
Action Identification in Gaussian Process Ban- information capacity; in particular, we can bound the maximum posterior
dits. icml 2021. [§b.4]
standard deviation after 𝑖 rounds of uncertainty sampling with:74
√︁
max 𝜎 = O( 𝛾𝑖 /𝑖). (10.43)
𝑥 ∈X

75 The pruning scheme ensures that (if the last By combining this result with the concentration inequality in (10.42),
batch concluded after 𝑖 steps) the global opti- and by periodically eliminating regions that are very likely to be subop-
mum is very likely among the remaining can-
didates, and thus the instantaneous regret in-
timal throughout optimization,75 the authors were able to show that the
curred in each stage
√︁ of the next batch will be
resulting algorithm has worst-case regret
bounded by O (𝛽𝑖 𝛾𝑖 /𝑖) with high probabil- √︁
ity. 𝑅¯𝜏 [𝐵] = O∗ ( 𝑑𝜏𝛾𝜏 ) (10.44)

76 m. valko et al. (2013). Finite-Time Analysis of with high probability, matching the regret bounds for gp-ucb and gp-ts
Kernelised Contextual Bandits. uai 2013. in the Bayesian setting (10.27).
77 s. salgia et al. (2020). A Computationally Ef- This is not the only result of this type; several other authors have
ficient Approach to Black-box Optimization been able to construct algorithms achieving similar worst-case regret
using Gaussian Process Models. arXiv: 2010.
13997 [stat.ML].
bounds through other means.76, 77, 78
78 r. camilleri et al. (2021). High-Dimensional
Experimental Design and Kernel Bandits. icml Lower bounds and tightness of existing algorithms
2021.
Lower bounds on regret are easier to come by in the frequentist setting
than in the Bayesian setting, as we have considerable freedom to con-
struct explicit objective functions that are provably difficult to optimize.
needles in haystacks: see p. 216 A common strategy to this end is to take inspiration from the “needle
in a haystack” trick discussed earlier. We construct a large set of suitably
79 j. scarlett et al. (2017). Lower Bounds on well-behaved “needles” that have little overlap, then argue that there
Regret for Noisy Gaussian Process Bandit Op- will always be some needle “missed” by an algorithm with insufficient
timization. colt 2017. budget to distinguish all the functions. Figure 10.5 shows a motivating
example with four translations of a smooth bump function with height
𝜀 and mutually disjoint support. Given any set of three observations –
regardless of how they were chosen – the cumulative regret for at least
one of these functions would be 3𝜀.
This construction embodies the spirit of most of the lower bound
The bump function used in scarlett et al.’s
arguments appearing in the literature. To yield a full proof, we must
lower bound analysis, the Fourier transform show how to construct a large number of suitable needles with bounded
of a smooth function with compact support rkhs norm. For a stationary process, this can usually be accomplished
as in figure 10.5. by scaling and translating a suitable bump-shaped function to cover the
domain. We also need to bound the regret of an algorithm given an input
80 This particular bump function is the prototyp- chosen from this set; here, we can keep the set of potential objectives
ical smooth function with compact support:
larger than the optimization budget and appeal to pigeonhole arguments.
(  
1
exp − 1−|𝑥 |𝑥 | < 1; In the frequentist setting with noise, the strongest known lower
𝑥 ↦→ |2
0 otherwise. bounds are due to scarlett et al.79 The function class considered in the
analysis was scaled and translated versions of a function similar to (in fact,

236
 10.6. the exact observation case

𝜀
Figure 10.5: Four smooth objective func-
tions with disjoint support.

precisely the Fourier transform of) the bump function in figure 10.5;80 this
function has the advantage of having “nearly compact” support while
having finite rkhs norm in the entire Matérn family. For optimization
on the unit cube X = [0, 1]𝑑 with the Matérn covariance function, the
authors were able to establish a lower bound on the cumulative regret
of any algorithm of

𝜈 +𝑑
𝑅¯𝜏 [𝐵] = Ω(𝜏 𝛼 ); 𝛼= ,
2𝜈 + 𝑑
and for the squared exponential covariance, a lower bound of
√

𝑅¯𝜏 [𝐵] = Ω 𝜏 (log 𝜏)𝑑 .

These bounds are fairly tight – they are within logarithmic factors of the
best-known upper bounds in the worst-case setting (10.15–10.16, 10.44);
see also the corresponding Bayesian bounds (10.28–10.29).
scarlett et al. also provided lower bounds on the worst-case simple
regret 𝑟¯𝜏 [𝐵], in terms of the expected time required to reach a given level
of regret. Inverting these bounds in terms of the simple regret at a given
time yields rates that are as expected in light of the relation in (10.4). 𝑟𝜏 ≤ 𝑅𝜏
𝜏

10.6 the exact observation case

The arguments outlined in the previous sections all ultimately depend on
the information capacity of noisy observations of a Gaussian process. An
upper bound on this quantity allows us to control the width of posterior
confidence intervals through a pivotal result (10.18). After noting that
the instantaneous regret of an upper confidence bound algorithm is in
turn bounded by the width of these confidence intervals (10.23), we may
piece together a bound on its cumulative regret (10.24).
Unfortunately, this line of attack breaks down with exact observa-
tions, as the information capacity of the now deterministic observation
process diverges to infinity.81 However, all is not lost – we can appeal to 81 This can be seen by taking 𝜎𝑛 → 0 in (10.17).
different techniques to find much stronger bounds on the width of con-
fidence intervals induced by exact observations, and thereby establish
much faster rates of convergence than in the noisy case.
Throughout this section, as in the previous two sections, we will assumptions: X is compact, 𝐾 (𝑥, 𝑥) ≤ 1 is
assume that the domain X ⊂ ℝ𝑑 is compact and that the covariance bounded
function is bounded by unity: 𝐾 (𝑥, 𝑥) ≤ 1.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 237

 theoretical analysis

Bounding the posterior standard deviation

As in the noisy setting, a key idea in deriving regret bounds with exact
maximum posterior standard deviation given observations is bounding the width of posterior confidence intervals
exact observations at x, 𝜎¯ x (10.39) resulting from observations. In this light, let us consider the
behavior of a Gaussian process GP (𝑓 ; 𝜇, 𝐾) on an objective 𝑓 : X → ℝ
82 We previously derived a bound of this form as it is conditioned on a set of exact observations. Given some set of
for uncertainty sampling in the noisy regime observation locations x ⊂ X, we seek to bound the maximum posterior
(10.43).
standard deviation of the process:
83 The literature on this topic is substantial, but
the following references provide good entry 𝜎¯x = max 𝜎,
points: 𝑥 ∈X

m. kanagawa et al. (2018). Gaussian Processes in terms of properties of x.82 This would serve as a universal bound on
and Kernel Methods: A Review on Connec- the width of all credible intervals, allowing us to bound the cumulative
tions and Equivalences. arXiv: 1807 . 02582 regret of a ucb-style algorithm via previous arguments.
[stat.ML]. [§ 5.2]
Thankfully, bounds of this type have enjoyed a great deal of atten-
h. wendland (2004). Scattered Data Approxi- tion, and strong results are available.83 Most of these results bound the
mation. Cambridge University Press. [chapter posterior standard deviation of a Gaussian process in terms of the fill
11]
distance, a measure of how densely a given set of observations fills the
domain. For a set of observation locations x ⊂ X, its fill distance is
fill distance of observation locations x, 𝛿 x defined to be
𝛿 x = max min |𝑥 − 𝑥𝑖 |, (10.45)
𝑥 ∈X 𝑥𝑖 ∈ x
the largest distance from a point in the domain to the closest observation.
Intuitively, we should expect the maximum posterior standard de-
viation induced by a set of observations to shrink with its fill distance.
This is indeed the case, and particularly nice results for the rate of this
shrinkage are available for the Matérn family. Namely, for finite 𝜈, once
the fill distance is sufficiently small (below a constant depending on the
covariance but not x), we have:

84 Carefully taking this limit yields the following 𝜎¯x ≤ 𝑐𝛿 x𝜈 , (10.46)

bound for the squared exponential covariance.
For sufficiently small 𝛿 x , we have where the constant 𝑐 does not depend on x. That is, once the observations
achieve some critical density, the posterior standard deviation of the
𝜎¯ x ≤ 𝛿 x𝑐/𝛿 x
process shrinks rapidly with the fill distance, especially as sample paths
for some 𝑐 not depending on x. Thus the rate become increasingly smooth in the limit 𝜈 → ∞.84This result will be
of convergence increases as 𝛿 x → 0:
instrumental in several results discussed below.
h. wendland (2004). Scattered Data Approxi-
mation. Cambridge University Press. [theorem
11.22] Optimizing fill distance
85 For an excellent survey on both maximin The question of optimizing the fill distance of a given number of points
(sphere packing) and minimax (fill distance)
optimal designs, see: in compact subsets of ℝ𝑑 – so as to maximally strengthen the bound on
prediction error above – is a major open question in geometry. Unfortu-
l. pronzato (2017). Minimax and maximin nately, this problem is exceptionally difficult due to its close connection
space-filling designs: some properties and
methods for construction. Journal de la Société to the notoriously difficult problem of sphere packing.85
Française de Statistique 158(1):7–36. In the unit cube X = [0, 1]𝑑, a simple grid strategy86 achieves asymp-
86 We successively exhaust all combinations totic fill distance O(𝜏 −1/𝑑 ), and thus for a Matérn covariance, we can
of dyadic rationals {𝑖/2𝑛 } with increasing guarantee (10.46):
height 𝑛 = 0, 1, . . . 𝜎¯x = O(𝜏 −𝜈/𝑑 ). (10.47)

238
 10.6. the exact observation case

This will suffice for the arguments to follow.87 However, this strategy may 87 We can extend this argument to any arbitrary
not be the best choice in practical settings, as the fill distance decreases compact domain X ⊂ ℝ𝑑 by enclosing it in a
cube.
sporadically by large jumps each time the grid is refined. An alternative
with superior “anytime” performance would be a low-discrepancy se-
quence such as a Sobol or Halton sequence, which would achieve slightly
larger (by a factor of log 𝜏), but more smoothly decreasing, asymptotic
fill distance.
Several sophisticated algorithms are also available for (approximately) 88 y. lyu et al. (2019). Efficient Batch Black-
optimizing the fill distance from a given fixed number of points.85, 88 box Optimization with Deterministic Regret
Bounds. arXiv: 1905.10041 [cs.LG].

Worst-case regret with deterministic observations

We now turn our attention to specific results, beginning with bounds on
the worst-case regret, where the analysis is somewhat simpler. As in the
noisy setting, we assume that the objective function lies in the rkhs ball worst-case regret with noise: p. 232
of radius 𝐵 corresponding to the covariance function 𝐾 used to model
the objective function during optimization, H𝐾 [𝐵].
In this setting, we have the remarkable result that posterior confi- 89 m. kanagawa et al. (2018). Gaussian Processes
dence intervals with fixed confidence parameter 𝐵 are always valid:89 and Kernel Methods: A Review on Connec-
tions and Equivalences. arXiv: 1807 . 02582
[stat.ML]. [corollary 3.11]
∀𝑥 ∈ X : 𝜙 ∈ [𝜇 − 𝐵𝜎, 𝜇 + 𝐵𝜎]. (10.48)

Combining this with the bound on standard deviation in (10.47) imme- 90 Here we use the alternative definition in foot-
diately shows that a simple grid strategy achieves worst-case simple note 5. As the confidence intervals (10.48) are
always valid, 𝑓 ∗ must always be in the (rapidly
regret90 shrinking) confidence interval of whatever
𝑟¯𝜏 [𝐵] = O(𝜏 −𝜈/𝑑 ) point maximizes the posterior mean.

for the Matérn covariance with smoothness 𝜈. bull provided a corre- 91 a. d. bull (2011). Convergence Rates of Effi-
sponding lower bound establishing that this rate is actually optimal: cient Global Optimization Algorithms. Jour-
nal of Machine Learning Research 12(88):2879–
𝑟¯𝜏 [𝐵] = Θ(𝜏 −𝜈/𝑑 ).91 As with previous results, the lower bound derives 2904.
from an adversarial “needle in haystack” construction as in figure 10.5.
This result is perhaps not as exciting as it could be, as it demonstrates convergence rates for expected improvement
that no adaptive algorithm can perform (asymptotically) better than grid
search in the worst case. However, we may reasonably seek similar
guarantees for algorithms that are also effective in practice. bull was
able to show that maximizing expected improvement yields worst-case
simple regret
𝑟¯𝜏 [𝐵] = O∗ (𝜏 − min(𝜈,1)/𝑑 ),
which is near optimal for 𝜈 ≤ 1. bull also showed that augmenting
expected improvement with occasional random exploration akin to an
𝜀-greedy policy improves its performance to near optimal for any finite
smoothness:
𝑟¯𝜏 [𝐵] = O∗ (𝜏 −𝜈/𝑑 ).
The added randomness effectively guarantees the fill distance of ob-
servations shrinks quickly enough that we can still rely on posterior
contraction arguments, and this strategy could be useful for analyzing
other policies.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 239

 theoretical analysis

Bayesian regret with deterministic observations

grünewälder et al. proved bounds on the Bayesian regret in the exact
observation case, although their results are only relevant for particularly
92 s. grünewälder et al. (2010). Regret Bounds rough objectives.92 The authors considered sample paths of a Gaussian
for Gaussian Process Bandit Problems. aistats process on the unit cube [0, 1]𝑑 with Lipschitz continuous mean function
2010.
and 𝛼-Hölder continuous covariance function. The latter is an exception-
ally weak smoothness assumption: the Matérn covariance with parame-
ter 𝜈 is Hölder continuous with 𝛼 = min(2𝜈, 1), and any distinction in
smoothness is lost beyond 𝜈 > 1/2, where sample paths are not yet even
differentiable. Under these assumptions, grünewälder et al. proved the
following bounds on the Bayesian simple regret, where the lower bound
is not universal but rather achieved by a specific construction satisfying
√︁
93 The upper bound is within a factor of log 𝜏 the assumptions:93
of wang et al.’s upper bound for the noisy √︁
√︁
𝛼
optimization of Brownian motion sample 𝔼[𝑟𝜏 ] = Ω 𝜏 −𝑐 / log 𝜏 ; 𝔼[𝑟𝜏 ] = O 𝜏 −𝑐 log 𝜏 ; 𝑐 = − ,
paths, where 𝛼 = 𝑑 = 1 (10.38). The lower 2𝑑
bound is identical to that case (10.37). The lower bound is achieved by an explicit Gaussian process with
Hölder continuous covariance function whose “needle in a haystack”
nature makes it difficult to optimize by any strategy. Specifically, the
authors show how to cover the unit cube with 2𝜏 disjoint and 𝛼-Hölder
continuous “needles” of compact support. We may then create a Gaussian
process by scaling each of these needles by an independent normal
random variable and summing. This construction is perhaps most clearly
demonstrated visually, and the marginal figure shows an example on
the unit interval with 𝜏 = 5 and 𝛼 = 1/2. As the needles are disjoint with
Samples from an example Gaussian process
used in deriving grünewälder et al.’s lower
independent heights, no policy with a budget of 𝜏 can determine more
bound. Here 10 ( 1/2)-Hölder continuous than half of the weights, and we must therefore pay a penalty in terms
“needles” are scaled and translated to cover of expected simple regret.
the unit interval, with one centered on each The upper bound derives from analyzing the performance of a simple
tick mark. Each needle is then scaled by an
independent normal random variable and
grid strategy via classical concentration inequalities.94 It is remarkable
summed, yielding a Gaussian process with that a nonadaptive strategy would yield such a small gap in regret com-
the desired properties. pared to the corresponding lower bound, but this result is probably best
understood as illustrating the inherent difficulty of optimizing rough
94 p. massart (2007). Concentration Inequalities functions rather than any inherent aptitude of grid search.
and Model Selection: Ecole d’Eté de Probabil- To underscore this remark, we may turn to a result of de freitas et al.,
ités de Saint-Flour xxxiii – 2003. Vol. 1896.
Springer–Verlag.
who showed that we may optimize sample paths of sufficiently smooth
Gaussian processes with exponentially decreasing expected simple regret
in the deterministic setting.95 This result relies on a few technical proper-
uniqueness of global maxima: § 2.7, p. 34 ties of the Gaussian process in question, in particular that it have a unique
global optimum and that it exhibit “nice” behavior in the neighborhood
95 n. de freitas et al. (2012b). Exponential Re- of the global optimum. A centered Gaussian process with covariance
gret Bounds for Gaussian Process Bandits with function in the Matérn family satisfies the required assumptions if 𝜈 > 2.
Deterministic Observations. icml 2012
The algorithm analyzed by the authors was a simple branch-and-
bound policy based on gp-ucb, wherein the domain is recursively subdi-
vided into finer and finer divisions. After each division, we identify the
regions that could still contain the global optimum based on the current
confidence intervals, then evaluate on these regions such that the fill
distance (10.45) is sufficiently small before the next round of subdivision.

240
 10.7. the effect of unknown hyperparameters

In this step, de freitas et al. relied on the bound in (10.46) (with 𝜈 = 2)

to ensure that the confidence intervals induced by observed data shrink
rapidly throughout this procedure.
At some point in this procedure, we will (with high probability) find
ourselves having rejected all but the local neighborhood of the global
optimum, at which point the assumed “nice” behavior of the sample
path guarantees rapid convergence thereafter. Specifically, the authors
demonstrate that at some point the instantaneous regret of this algorithm
will converge exponentially:
  
𝑐𝜏
𝔼[𝜌𝜏 ] = O exp − ,
(log 𝜏)𝑑/4

for some constant 𝑐 > 0 depending on the process but not on 𝜏. This 96 This is not inconsistent with grünewälder et
condition implies rapid convergence in terms of simple regret as well al.’s lower bound: here we assume smoothness
consistent with 𝜈 > 2 in the Matérn family,
(as we obviously have 𝑟𝜏 ≤ 𝜌𝜏 ) and bounded cumulative regret after we whereas the adversarial Gaussian process con-
have entered the converged regime.96 structed in the lower bound is only as smooth
Evidently optimization is much easier with deterministic observa- as 𝜈 = 1/2.
tions than noisy ones, at least in terms of Bayesian regret. Intuitively,
the reason for this discrepancy is that noisy observations may compel
us to make repeated measurements in the same region in order to shore
up our understanding of the objective function, whereas this is never
necessary with exact observations.

10.7 the effect of unknown hyperparameters

We have now outlined a plethora of convergence results: upper bounds on
the regret of specific algorithms and algorithm-agnostic lower bounds,
in the expected and worst case, with and without observation noise.
However, all of these results assumed intimate knowledge about the
objective function being optimized: in Bayesian analysis, we assumed
the objective is sampled from the prior used to model it, and for worst-
case analysis, we assumed the objective lay in the corresponding rkhs.
Of course, neither of these assumptions (especially the former!) may
hold in practice. Further, in practice the model used to reason about
the objective is typically inferred from observed data and constantly
updated throughout optimization, but the analysis discussed thus far has modeling with Gaussian processes: § 3, p. 45
assumed the model is not only perfectly informed, but also fixed.
It turns out that violations to these (rather implausible!) assumptions
can be disastrous for convergence. For example, consider this prototypi-
cal Bayesian optimization algorithm: we model the objective function
with an automatic relevance determination (ard) version of a Matérn
covariance function, learn its output (3.20) and length scales (3.25) via automatic relevance determination: § 3.4, p. 56
maximum likelihood estimation, and design observation locations by ml estimation of hyperparameters: § 4.3, p. 73
maximizing expected improvement. Although this setup has proven
remarkably effective in practice,97 bull proved that it can actually fail 97 j. snoek et al. (2012). Practical Bayesian Op-
miserably in the frequentist setting: we may construct functions in the timization of Machine Learning Algorithms.
neurips 2012.
rkhs of any such covariance function (that is, with any desired param-

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 241

 theoretical analysis

Figure 10.6: The problem with estimat-

ing hyperparameters. Above:
a function in the rkhs for
the 𝜈 = 5/2 Matérn covari-
ance with unit norm. Be-
low: a sample path from
a Gaussian process with
the same covariance. Unless
we’re lucky, we may never
find the “hump” in the for-
mer function and thus may
never build a reasonable be-
lief regarding the objective.
eters) that will “fool” this algorithm into having high regret with high
98 a. d. bull (2011). Convergence Rates of Effi- probability.98
cient Global Optimization Algorithms. Jour- Estimation of hyperparameters can also cause problems with con-
nal of Machine Learning Research 12(88):2879–
2904. [theorem 3]
vergence in the Bayesian setting. For example, locatelli considered
optimization on the unit interval X = [0, 1] from exact observations
99 m. locatelli (1997). Bayesian Algorithms for using a Wiener process prior and maximizing expected improvement. In
One-Dimensional Global Optimization. Jour- this relatively straightforward model, the only hyperparameter under
nal of Global Optimization 10(1):57–76.
consideration was an output scale (3.20). locatelli showed that for a
fixed output scale, this procedure converges for all continuous func-
tions.99 However, when the output scale is learned from data – even in a
fully Bayesian manner where it is endowed with a prior distribution and
marginalized in the predictive distribution – there are extremely simple
(piecewise linear!) functions for which this algorithm does not recover
the optimum. An example is shown in the margin.
1 3 2 In both of these examples, the roadblock to convergence when esti-
mating hyperparameters is a mismatch between the objective function
A function given by locatelli for which
expected improvement combined with model used by the Bayesian optimization algorithm and the true objec-
a Wiener process prior will not con- tive. In the frequentist setting, there is an inherent tension as functions
verge if its output scale is marginalized. lying in the rkhs H𝐾 are not representative of sample paths from the
The first three observations are at the
indicated locations; after iteration 3, the
Gaussian process GP (𝑓 ; 𝜇, 𝐾).100
left hand side of the domain is forever ignored. Functions in the rkhs are much smoother than sample paths and may
feature relatively “flat” regions of little variation, which may in turn lead
100 Recall that sample paths almost surely do not to poorly fit models. See figure 10.6 for a striking example. locatelli’s
lie in H𝐾 , assuming it is infinite dimensional, piecewise linear counterexample also reflects extreme model misspeci-
and see figure 10.2 and the surrounding text. fication – the function is far too smooth to resemble Brownian motion
in any meaningful statistical sense, and we should not be surprised that
maximum likelihood estimation leads to an inconsistent algorithm.
A line of continuing work has been to derive robust algorithms that
do not require perfect prior knowledge regarding the objective function
in order to provide strong convergence guarantees. There are several
ideas in this direction, all of which employ some mechanism to ensure
that the space of objective functions considered by the algorithm does
101 bull simply inflated the output scale esti- not ever become too “small.” For example, both bull and locatelli
mated
√ by maximum likelihood by a factor of addressed their counterexamples with schemes wherein the output scale
𝜏, and locatelli chose a prior on the output
scale that had no support below some mini- used during optimization is never allowed to shrink so rapidly that true
mum threshold. objective function is left behind.101

242
 10.8. summary of major ideas

In the frequentist setting, one common strategy is to replace the

assumption that the objective function lay in some particular rkhs with
the assumption that it lay in some parametric family of rkhses indexed
by a set of hyperparameters. We then slowly expand the space of func-
tions considered in the algorithm over the course of the algorithm such
that the objective function is guaranteed to eventually – and forever
thereafter – be well explained. In particular, consider augmenting an
isotropic covariance function from the Matérn family 𝐾 with an output
scale (3.20) 𝜆 and a vector of length scales ℓ (3.25). We have the remark-
able property that if a function 𝑓 is in the rkhs ball of radius 𝐵 (10.11)
for some setting of these hyperparameters:

𝑓 ∈ H𝐾 [𝐵; 𝜆, ℓ],

then the same function is also in the rkhs ball for any larger output
scale and for any vector of shorter (as in the lexicographic order) length
scales:
𝑓 ∈ H𝐾 [𝐵; 𝜆,0 ℓ 0]; 𝜆 0 ≥ 𝜆; ℓ 0 ≤ ℓ.
With this in mind, a natural idea for deriving theoretical convergence 102 z. wang and n. de freitas (2014). Theoretical
results (but not necessarily for realizing a practical algorithm!) is to Analysis of Bayesian Optimization with Un-
known Gaussian Process Hyper-Parameters.
ignore any data-dependent scheme for setting hyperparameters and arXiv: 1406.7758 [stat.ML].
instead simply slowly increase the output scale and slowly decrease the
103 f. berkenkamp et al. (2019). No-Regret Bayes-
length scales over the course of optimization, so that at some point any ian Optimization with Unknown Hyperparam-
function lying in any such rkhs will eventually be captured. This scheme eters. Journal of Machine Learning Research
has been used to provide convergence guarantees for both expected 20(50):1–24.
improvement102 and gp-ucb103 with unknown hyperparameters.

10.8 summary of major ideas

• Convergence analysis for Bayesian optimization algorithms entails se-
lecting a measure of optimization performance, a space of objective
functions to consider, and deriving bounds on the asymptotic growth of
the chosen performance measure on the chosen function space.
• Optimization performance is almost always assessed via one of two re- simple and cumulative regret: § 10.1, p. 213
lated notations of regret, both of which depend on the difference between
the function value at measured locations and the global optimum.
• Although tempting, the space of all continuous functions is too large to useful function spaces for studying
be of much interest in analysis: we may construct “needles in haystacks” convergence: § 10.2, p. 215
to foil any algorithm by any amount. Instead, we may study spaces of
objective functions with more plausible behavior. A Gaussian process
model suggests two natural possibilities: sample paths from the process
and the reproducing kernel Hilbert space (rkhs) associated with the
process, the closure of the space of all possible posterior mean functions.
• Putting these pieces together, a Bayesian analysis of regret assumes the upper bounds:
objective is a sample path from a Gaussian process and seeks asymptotic Bayesian regret with noise: § 10.4, p. 224
bounds on the expected regret (10.6). A frequentist analysis, on the other worst-case regret with noise: § 10.5, p. 232
hand, assumes the objective function is a fixed, but unknown function Bayesian regret without noise: § 10.6, p. 240

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 243

theoretical analysis

worst-case regret without noise: § 10.6, p.239 in a given rkhs and seeks asymptotic bounds on the worst-case regret
(10.6).
this argument is carefully laid out for • Most upper bounds on cumulative regret derive from a proof strategy
Bayesian regret with noise in § 10.4, p.225 where we identify some suitable set of predictive credible intervals that
are universally valid with high probability. We may then argue that
the cumulative regret is bounded in terms of the total width of these
intervals. This strategy lends itself most naturally to analyzing the Gaus-
sian process upper confidence bound (gp-ucb) policy, but also yields
bounds for Gaussian process Thompson sampling (gp-ts) due to strong
theoretical connections between these algorithms (10.33).
information capacity: § 10.3, p. 222 • In the presence of noise, a key quantity is the information capacity, the
maximum information about the objective that can be revealed by a set
of noisy observations. Bounds on this quantity yield bounds on the sum
bounding the posterior standard deviation of predictive variances (10.18) and thus cumulative regret. With exact
with exact observations: § 10.6, p. 238 observations, we may derive bounds on credible intervals by relating
the fill distance (10.45) of the observations to the maximum standard
deviation of the process, as in (10.46).
lower bounds: • To derive lower bounds on Bayesian regret, we seek arguments limiting
Bayesian regret with noise: § 10.4, p. 230 the rate at which Gaussian process sample paths can be optimized. For
worst-case regret with noise, § 10.5, p. 236 worst-case regret, we may construct explicit objective functions in a
Bayesian regret without noise: § 10.6, p. 240 given rkhs and prove that they are difficult to optimize; here the “needles
worst-case regret without noise: § 10.6, p.239 in haystacks” idea again proves useful.

244
 11
EXTENSIONS AND RELATED SET TINGS

Thus far we have focused exclusively on a simple model problem (al-

gorithm 1.1): sequential optimization of a single objective with either a
fixed evaluation budget or known observation costs. These assumptions
are convenient for study and often reasonable in practice, but not all
optimization scenarios fit neatly into this mold. Numerous extensions of
this setup have received serious attention in the literature, and we will
provide an overview of the most important of these in this chapter.
A running theme throughout this discussion will be adapting the
decision-theoretic framework developed in chapter 5 to derive policies decision theory for optimization: chapter 5,
for each of these new settings. In that chapter, we derived optimal opti- p. 87
mization policies for our model problem, but we want to stress that the
overarching approach to decision making can be extended to effectively
any scenario.
Namely, we may derive an optimal policy for any sequential experi- general procedure for optimal policies
mental design problem by following a simple recipe, an abstraction of
our previous presentation:
1. Identify the action space A of each decision.
2. Define preferences over outcomes with a utility function 𝑢 (D).
3. Identify the uncertain elements 𝜓 relevant for each decision and deter-
mine how to compute the posterior belief given data, 𝑝 (𝜓 | D).
4. Compute the one-step marginal gain 𝛼 1 (5.8) for every action.1 1 Conveniently, this step also yields the one-step
lookahead approximate policy as a side effect.
5. Derive the optimal policy by induction on the horizon (5.15–5.16).
We will sketch how this scheme can be realized for notable optimization
settings not yet addressed. Each will involve additional complexity in
building effective policies due to additional complexity in the problem
formulation, but we will demonstrate how we can adapt our existing
tools by addressing the appropriate steps of the above general procedure.
We will also provide a survey of proposed approaches to each of the
optimization extensions we will consider, regardless of whether they are
grounded in decision theory.

11.1 unknown observation costs

In our discussion on cost-aware optimization, we assumed that observa- cost-aware optimization: § 5.4, p. 103
tion costs were prescribed by a known cost function 𝑐, which enabled cost function, 𝑐
a simple mechanism for dynamic termination through careful account-
ing. However, in some scenarios the cost of an observation may not
be known a priori, but rather only determined through the course of
acquisition. These unknown costs reflect additional uncertainty that
must be accounted for in policy design, as the utility of the returned data
depends critically on their values. The natural solution is to perform
inference about observation costs throughout optimization and use the
resulting beliefs to guide our decisions, just as we do with an unknown

This material will be published by Cambridge University Press as Bayesian Optimization. 245
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 extensions and related settings

objective function. To adapt the cost-aware policies from chapter 5 to

this setting, we must revisit steps 3–4 of the above procedure.

Inference for unknown cost function

At a high level, reasoning about an unknown cost function given ob-
servations is no different from reasoning about an unknown objective
function, and we can apply any suitable regression model for this task.
The details of this inference will obviously depend on the situation, but
we can outline one rather general approach.
notation for observation costs First let us define some notation for cost observations mirroring
our notation for the objective function. Suppose that evaluation costs
are determined, perhaps stochastically, by an underlying cost function
cost function, 𝑐 𝑐 : X → ℝ we wish to infer. Suppose further that an evaluation at a point
𝑥 now returns both a measured value 𝑦, whose distribution depends on
the corresponding objective function value 𝜙 = 𝑓 (𝑥), and an observation
observed cost at 𝑥, 𝑧 cost 𝑧, whose distribution depends on the corresponding cost function
value of cost function at 𝑥, 𝜅 = 𝑐 (𝑥) value 𝜅 = 𝑐 (𝑥). We will accumulate these values throughout optimization
in an augmented dataset of observations and their costs, D = (x, y, z).
If we can safely model cost observations as conditionally indepen-
dent of objective observations given the chosen location, then we may
follow our modeling strategy for the objective function and assume that
cost observation model, 𝑝 (𝑧 | 𝑥, 𝜅) each observed cost is generated by an observation model 𝑝 (𝑧 | 𝑥, 𝜅).
This allows for modeling a wide range of different scenarios, including
nondeterministic costs.2 Now we can proceed with inference about the
cost function as usual. We choose a suitable (possibly parametric) prior
cost function prior, 𝑝 (𝑐) process 𝑝 (𝑐), which we condition on observed costs to form the posterior
cost function posterior, 𝑝 (𝑐 | D) 𝑝 (𝑐 | D). Finally, we can form the predictive distribution for the cost
of making an observation at an arbitrary point 𝑥 by marginalizing the
predictive distribution for cost, 𝑝 (𝑧 | 𝑥, D) latent cost function:
∫
𝑝 (𝑧 | 𝑥, D) = 𝑝 (𝑧 | 𝑥, 𝜅) 𝑝 (𝜅 | 𝑥, D) d𝜅.

2 As an example use case, suppose evaluating In some applications, observation costs may be nontrivially corre-
the objective requires train a machine learning lated with the objective function. As an extreme example, consider a
model on cloud hardware with variable (e.g.,
spot) pricing.
common problem in algorithm configuration,3 where the goal is to design
the parameters of an algorithm so as to minimize its expected running
3 The related literature is substantial, but a
Bayesian optimization perspective can be
time. Here the cost of evaluating a proposed configuration might be
found in: reasonably defined to be proportional to its running time. Up to scaling,
the observation cost is precisely equal to the objective! To model such
f. hutter et al. (2011). Sequential Model-Based
Optimization for General Algorithm Configu- correlations, we could define a joint prior 𝑝 (𝑓, 𝑐) over the cost and ob-
ration. lion 5. jective functions, as well as a joint observation model 𝑝 (𝑦, 𝑧 | 𝑥, 𝜙, 𝜅).
We could then continue as normal, computing expected utilities with
respect to the joint predictive distribution
∬
𝑝 (𝑦, 𝑧 | 𝑥, D) = 𝑝 (𝑦, 𝑧 | 𝑥, 𝜙, 𝜅) 𝑝 (𝜙, 𝜅 | 𝑥, D) d𝜙 d𝜅,

a setup offering considerable flexibility in modeling.

246
 11.1. unknown observation costs

observations posterior mean posterior 95% credible interval

cost-agnostic expected improvement, 𝛼 ei

posterior mean, cost
posterior 95% credible interval, cost

cost-adjusted expected improvement next observation location

Figure 11.1: Decision making with uncertain costs. The middle panel shows the cost-agnostic expected improvement
acquisition function along with a belief about an uncertain cost function, here assumed to be independent of
the objective. The bottom panel shows the cost-adjusted expected improvement, marginalizing uncertainty in
the objective and cost function (11.1).

Decision making with unknown costs

The approach outlined above suffices to maintain a belief about the poten-
tial cost of observations proposed throughout optimization, but we still
must account for this uncertainty in the optimization policy. Thankfully,
this is relatively straightforward in our decision-theoretic framework: we
simply compute expected utility accounting for all relevant uncertainty
as usual, here to include cost. Given an arbitrary dataset D, now aug-
mented with observation costs, consider the one-step expected marginal
gain in utility:
 
𝛼 1 (𝑥; D) = 𝔼 𝑢 (D1 ) | 𝑥, D − 𝑢 (D).

Computing this expectation now requires integrating over both the

unknown measurement 𝑦 and the unknown observation cost 𝑧:
∬
 

𝔼 𝑢 (D1 ) | 𝑥, D = 𝑢 D ∪ {𝑥, 𝑦, 𝑧} 𝑝 (𝑦, 𝑧 | 𝑥, D) d𝑦 d𝑧. (11.1)
4 Of course, all nested expectations must also be
Although there is some slight added complexity in this computation, taken with respect to unknown observation
the optimal policy otherwise remains exactly as derived in (5.15–5.17).4 costs!

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 247

 extensions and related settings

observations posterior mean posterior 95% credible interval

cost-agnostic expected improvement, 𝛼 ei

posterior mean, cost
posterior 95% credible interval, cost

expected improvement per unit cost next observation location

Figure 11.2: Expected gain per unit cost. The middle panel shows the cost-agnostic expected improvement acquisition
function along with a belief about an uncertain cost function, here assumed to be independent of the objective.
The bottom panel shows the expected improvement per unit cost (11.1).

Figure 11.1 illustrates this policy, combining expected improvement with

an independent uncertain cost function.

Expected gain per unit cost

Other approaches to dealing with unknown costs are also of course
possible. snoek et al. proposed one notable option that has gained some
5 s. zilberstein (1996). Using Anytime Algo- popularity based on a heuristic common in anytime algorithms5 – that is,
rithms in Intelligent Systems. ai Magazine algorithms that seek to maximize the instantaneous rate of improvement
17(3):73–83.
under the premise that the procedure may be terminated at any time.6
6 j. snoek et al. (2012). Practical Bayesian Op- The idea is simple: we can approximate the expected immediate marginal
timization of Machine Learning Algorithms.
neurips 2012. gain in utility per unit cost from an observation at 𝑥 by
 
𝑢 (D1 ) − 𝑢 (D)  𝛼 1 (𝑥; D)
𝔼 𝑥, D ≈ , (11.2)
𝜅 𝔼[𝜅 | 𝑥, D]
where 𝛼 1 is the expected gain in data utility. This is a first-order approx-
imation to the expected gain-per-cost (which is not the ratio of their
7 Assuming independence between the objec-
tive and cost, a second-order expansion is just
respective expectations, even in the independent case) that could be
as easy to compute: further refined if desired,7 but works well in practice. The motivating ex-
𝛼 1 (𝑥; D) 𝛼 1 (𝑥; D) var[𝜅 | 𝑥, D]
ample for snoek et al. was hyperparameter tuning of machine-learning
𝔼[𝜅 | 𝑥, D]
+
𝔼[𝜅 | 𝑥, D] 3
. algorithms with unknown training costs, and the simple heuristic of
maximizing “expected improvement per (expected) second” delivered

248
 11.2. constrained optimization and unknown constraints

promising results in their experiments. This heuristic has since appeared

in other contexts.8 8 g. malkomes et al. (2016). Bayesian optimiza-
tion for automated model selection. neurips
Figure 11.2 illustrates this policy with the sample example in figure 2016.
11.2. The chosen decision closely matches the decision reached in figure
11.1. It is interesting to compare the behavior of the two acquisition
functions on both sides of the domain: whereas these regions are not
especially exciting in the additive cost approach, they are appealing from
the anytime view – although they are expected to give only modest
improvement, they are also relatively inexpensive.

11.2 constrained optimization and unknown constraints

Many optimization problems feature constraints restricting allowable
solutions in a potentially complex, even uncertain manner. To this end
we may extend our running optimization problem (1.1) to incorporate
arbitrary constraints; a common formulation is:

𝑥 ∗ ∈ arg max 𝑓 (𝑥) subject to ∀𝑖 : 𝑔𝑖 (𝑥) ≤ 0,

𝑥 ∈X

where the functions {𝑔𝑖 } : X → ℝ comprise a set of inequality con- inequality constraints, {𝑔𝑖 }
straints. The subset of the domain where all constraints are satisfied is
known as the feasible region: feasible region, F

F = 𝑥 ∈ X | ∀𝑖 : 𝑔𝑖 (𝑥) ≤ 0 .

In some situations, the value of some or all of the constraint functions uncertain constraints
may in fact be unknown a priori and only revealed through experimenta-
tion, complicating policy design considerably. As an example, consider a
business optimizing the parameters of a service to maximize revenue,
subject to constraints on customer response. If customer response is
measured experimentally – for example via a focus group – we cannot
know the feasibility of a proposed solution until after the objective has
been measured, and even then only with limited confidence.
Further, even if the constraint functions can be computed exactly on
demand, constrained optimization of an uncertain objective function is
not entirely straightforward. In particular, an observation of the objective
at an infeasible point may yield useful information regarding behavior
on the feasible region, and could represent an optimal decision if that
information were compelling enough. Thus simply restricting the action
space to the feasible region may not be the best approach to policy 9 r. b. gramacy and h. k. h. lee (2011). Optimiza-
design.9 Instead, to derive effective policies for constrained optimization, tion Under Unknown Constraints. In: Bayesian
we must reconsider steps 2–4 of our general approach.
Statistics 9.

Modeling constraint functions

To allow for uncertain constraint functions, we begin by modeling the
joint observation process of the objective and constraint functions. As
with an uncertain cost function, we will assume that each observation of

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 249

 extensions and related settings

the objective is accompanied by some information regarding constraint

satisfaction at the chosen location. Modeling this process is trivial when
the constraint functions are known a priori, but otherwise may require
some care.
It is difficult to provide concrete advice as information about con-
straint satisfaction may assume different forms, ranging from exact
observation of the constraint functions to mere binary indicators of
constraint satisfaction. In some situations, we may even face stochastic
constraint processes where the feasibility of a given location is only
achieved with some unknown probability. Fortunately, our discussion in
the previous section regarding modeling an uncertain cost function is
general enough to handle any of these situations by choosing an appro-
priate joint prior processes and observation model for the objective and
constraint functions.

Defining a utility function

Next we must define a utility function appropriate for constrained op-
timization. Most of the utility functions in chapter 6 can be suitably
modified to this end.
simple reward: § 6.1, p. 112 We can realize a constrained version of simple reward by considering
the expected utility of a risk-neutral terminal recommendation following
optimization. As before, the resulting utility will depend on the action
space used for the terminal decision. Perhaps the simplest option would
be to limit the recommendation to the feasible observed locations (if
known!):
𝑢 (D) = max 𝜇D (𝑥), (11.3)
𝑥 ∈ x∩F
resulting in a natural adaptation of the simple reward (6.3). If the entire
feasible region is known, we might instead allow recommending any
feasible point, giving rise to an adaptation of the global simple reward
(6.5):
𝑢 (D) = max 𝜇D (𝑥). (11.4)
𝑥 ∈F
Finally, in the case of uncertain constraints, we might limit our rec-
10 m. a. gelbart et al. (2014). Bayesian Optimiza- ommendation to those points believed to be feasible with sufficient
tion with Unknown Constraints. uai 2014. confidence:10

𝑢 (D) = max 𝜇D (𝑥); F (𝛿) = 𝑥 | Pr(𝑥 ∈ F | D) ≥ 1 − 𝛿 . (11.5)
𝑥 ∈F (𝛿)

With some care, we could also modify other utility functions to be

aware of a (possibly uncertain) feasible region, although the variations
on simple reward above have received the most attention.

Deriving a policy
After selecting a model for the constraint functions and a utility function
for our observations, we can derive a policy for constrained optimization
following the standard procedure of induction on the decision horizon.

250
 11.2. constrained optimization and unknown constraints

A policy that has received particular attention is the result of one-step

lookahead with (11.3).10, 11, 12 If we assume that the constraint functions are 11 m. schonlau et al. (1998). Global versus Local
conditionally independent of the objective function given the observation Search in Constrained Optimization of Com-
puter Models. In: New Developments and Ap-
location, then the one-step expected marginal gain in utility becomes plications in Experimental Design.
0
𝛼 (𝑥; D) = 𝛼 ei (𝑥; D, 𝜇 ∗ ) Pr(𝑥 ∈ F | 𝑥, D). (11.6) 12 j. r. gardner et al. (2014). Bayesian Optimiza-
tion with Inequality Constraints. icml 2014.
This is simply the expected improvement, measured with respect to the
feasible incumbent value 𝜇 ∗ = 𝑢 (D) (7.21, 11.3), weighted by the probabil-
ity of feasibility, a natural policy we might arrive at via purely heuristic
arguments. gelbart et al. point out this acquisition function has a slight
pathology (also present with unconstrained expected improvement): the
utility degenerates when no feasible observations are available, and (11.6)
becomes ill-defined. In this case, the authors propose simply maximizing
the probability of feasibility:

𝛼 (𝑥; D) = Pr(𝑥 ∈ F | 𝑥, D).

The expected feasible improvement (11.6) encodes a strong prefer- observations in the infeasible region
ence for evaluating on the feasible region only, and in the case where
the constraint functions are all known, the resulting policy will never
evaluate outside the feasible region.13 This is a natural consequence of 13 In this case the policy would be equivalent to
the one-step nature of the acquisition function: an infeasible observation redefining the domain to encompass only the
feasible region F and maximizing the unmod-
cannot yield any immediate improvement to the utility (11.3) and thus ified expected improvement (7.2).
cannot be one-step optimal. However, this behavior might be seen as un-
desirable given our previous comment that infeasible observations may
yield valuable information about the objective on the feasible region.
If we wish to realize a policy more open to observing outside the
feasible region, there are several paths forward. A less-myopic policy
built on the same utility (11.3) is one option; even two-step lookahead
could elect to obtain an infeasible measurement. Another possibility is
one-step lookahead with a more broadly defined utility such as (11.4–11.5),
which can see the merit of infeasible observations through more global
evaluation of success.
To encourage infeasible observations when prudent, gramacy and 14 r. b. gramacy and h. k. h. lee (2011). Optimiza-
lee proposed a score they called the integrated expected conditional tion Under Unknown Constraints. In: Bayesian
Statistics 9.
improvement:14
∬
 
𝛼 ei (𝑥 0; D) − 𝛼 ei (𝑥 0; D 0) Pr(𝑥 0 ∈ F | 𝑥 ,0 D) 𝑝 (𝑦 | 𝑥, D) d𝑥 0d𝑦,

where D 0 is the putative updated dataset and the location 𝑥 0 is integrated

over the domain. This is a measure of the expected impact of a measure-
ment on the entire acquisition surface over the feasible region, which
can effectively capture the potential impact of an infeasible observation
when it is useful. A similar approach was taken by picheny, who inte-
grated the change in probability of improvement against the feasibility 15 v. picheny (2014). A Stepwise uncertainty re-
probability.15 Although these approaches can be heuristically motivated, duction approach to constrained global opti-
mization. aistats 2014.
the required integrals over the acquisition surfaces are intractable, and
no obvious approximations are available beyond standard methods.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 251

 extensions and related settings

decoupled observations gelbart et al. considered a variation on the constrained optimiza-

tion problem discussed above wherein observations of the objective and
16 m. a. gelbart et al. (2014). Bayesian Optimiza- constraints can be “decoupled” – that is, when we can elect to measure
tion with Unknown Constraints. uai 2014. any of these functions independent of the others, expanding the action
17 j. m. hernández-lobato et al. (2016b). A Gen- space of the decision problem.16 The authors noted the expected feasi-
eral Framework for Constrained Bayesian Op- ble improvement (11.6) displayed undesirable behavior in this scenario
timization using Information-based Search.
Journal of Machine Learning Research 17:1–53.
and proposed an alternative policy based on mutual information, which
was later refined and expanded into a fully fledged information theo-
retic policy for constrained optimization (with or without decoupled
predictive entropy search: § 8.8, p. 180 observations) based on predictive entropy search.17

11.3 synchronous batch observations

Many optimization settings allow for the possibility of making multiple
observations in parallel. In fact, some settings such as high-throughput
screening for scientific discovery practically demand parallel experiments
due to the growing capacity of sophisticated automated instruments.
Numerous batch policies have been proposed for Bayesian optimization
to harness this capability, including variants of virtually every popular
sequential policy.
synchronous vs. asynchronous batch Here we can distinguish two settings: synchronous and asynchronous
construction batch construction. In both cases, multiple experiments must be designed
to run in parallel. The distinguishing factor is that in the synchronous
case, the results from each entire batch of experiments are obtained
before designing the next, whereas in the asynchronous case, each time
an experiment completes, we may immediately design a new one in
light of those still pending. Bayesian decision theory naturally offers
one possible approach to both of these scenarios. We will focus on the
asynchronous batch observations: § 11.4, synchronous case in this section and will discuss the asynchronous case
p. 262 in the next section.

Decision-theoretic batch construction

Consider an optimization scenario where in each iteration we may de-
batch of observation locations, x sign a batch of 𝑏 points x = {𝑥 1, 𝑥 2, . . . 𝑥𝑏 } for simultaneous evaluation,
corresponding observed values, y resulting in a corresponding vector of measured values y obtained before
our next action. The design of each batch represents a decision with
action space for batch observations, A = X 𝑏 action space A = X 𝑏, a modification to step 3 of our general procedure.
We proceed by computing the one-step expected gain in utility from
expected one-step marginal gain from batch a proposed batch measurement x. We will call the corresponding batch
observation, 𝛽 1 acquisition function 𝛽 1 to distinguish it from its sequential analog 𝛼 1 :
 
𝛽 1 (x; D) = 𝔼 𝑢 (D1 ) | x, D − 𝑢 (D).
Here D1 representsthe data available after the batch observation is
resolved: D1 = D ∪ (x, y) . Computing this expected marginal gain is
an expectation with respect to the unknown values y:
∫
𝛽 1 (x; D) + 𝑢 (D) = 𝑢 (D1 ) 𝑝 (y | x, D) dy, (11.7)

252
 11.3. synchronous batch observations

observations
posterior mean
posterior 95% credible interval

batch expected improvement

optimal batch 𝑥0
𝑥

𝑥
𝑥0

0 ∗
𝛽 ei

Figure 11.3: Optimal batch selection. The heatmap shows the expected one-step marginal gain in simple reward (6.1) from
adding a batch of two points – corresponding in location to the belief about the objective plotted along the
margins – to the current dataset. Note that the expected marginal gain is symmetric. The optimal batch will
observe on both sides of the previously best-seen point. In this example, incorporating either one of the
selected points alone would also yield relatively high expected marginal gain.

and an optimal batch with decision horizon 𝜏 = 1 maximizes this score:

x ∈ arg max 𝛽 1 (x0; D). (11.8)
x0 ∈X 𝑏

Finally, we can derive the optimal batch policy for a fixed evaluation optimal batch policy with fixed evaluation
budget by induction on the horizon, accounting for the expanded action budget
space for each future decision:
 ∗ 
x ∈ arg max 𝛽 1 (x0; D) + 𝔼 𝛽𝜏−1 (D1 ) | x,0 D ;
x0 ∈X 𝑏
=𝛽𝜏 (x0;D)

𝛽𝜏∗ (D) = max 𝛽𝜏 (x0; D).

x0 ∈X 𝑏

If desired, we could also allow for variable-cost observations and the variable costs and termination option
option of dynamic termination by accounting for costs and including a
termination option in the action space. Another compelling possibility
would be to consider dynamic batch sizes by expanding the action space dynamic batch sizes
further and assigning an appropriate size-dependent cost function for
proposed batch observations.
Optimal batch selection is illustrated in figure 11.3 for designing a example of optimal batch policy
batch of two points with horizon 𝜏 = 1. We compute the expected one-
step gain in utility (11.7) – here the simple reward (6.1), analogous to

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 253

 extensions and related settings

expected improvement 𝛼 ei (7.2) – for every possible batch and observe

where the score is maximized. The optimal batch evaluates on either
side of the previously best-seen point, achieving distributed exploitation.
The expected marginal gain surface has notably complex structure, for
example expressing a strong preference for batches containing at least
one of the chosen locations over any purely exploratory alternative, as
well as severely punishing batches containing an observation too close
to an existing one.
connection to 𝑏-step lookahead We may gain some insight into the optimal batch policy by decom-
posing the expected batch marginal gain in terms of corresponding
quantities from the optimal sequential policy. Let us first consider the
expected marginal gain from selecting a batch of two points, x = {𝑥, 𝑥 0 },
resulting in observed values y = {𝑦, 𝑦 0 }. Let D 0 represent the current
data augmented with the single observation (𝑥, 𝑦). We may rewrite the
marginal gain from the batch observation (x, y) as a telescoping sum
with terms corresponding to the impact of each individual observation:
   
𝑢 (D1 ) − 𝑢 (D) = 𝑢 (D1 ) − 𝑢 (D 0) + 𝑢 (D 0) − 𝑢 (D) ,

which allows us to rewrite the one-step expected batch marginal gain as:
 
𝛽 1 (x; D) = 𝛼 1 (𝑥; D) + 𝔼 𝛼 1 (𝑥 0; D 0) | x, D .

This expression is remarkably similar to the optimal two-step expected

sequential marginal gain (5.12):
h i
𝛼 2 (𝑥; D) = 𝛼 1 (𝑥; D) + 𝔼 max
0
𝛼 1 (𝑥 0
; D 0
) | 𝑥, D .
𝑥 ∈X

The main difference is that in the batch setting, we must commit to both
observation locations a priori, whereas in the sequential setting, we can
design our second observation optimally given the outcome of the first.
“unrolling” the optimal policy: § 5.3, p. 99 We can extend this relationship to the general case. Temporarily
adopting compact notation, a horizon-𝑏 optimal sequential decision
satisfies:
n   o
𝑥 ∈ arg max 𝛼 1 + 𝔼 max 𝛼 1 + 𝔼[max{𝛼 1 + · · · , (11.9)

and the optimal one-step batch of size 𝑏 satisfies:

n  o
x ∈ arg max 𝛼 1 + 𝔼 𝛼 1 + 𝔼[ 𝛼1 + · · · . (11.10)

Clearly the expected utility gained from making 𝑏 optimal sequential

18 j. vondrák (2005). Probabilistic Methods in decisions surpasses the expected utility from a single optimal batch the
Combinatorial and Stochastic Optimization. same size: the sequential policy benefits from designing each successive
Ph.D. thesis. Massachusetts Institute of Tech-
nology.
observation location adaptively, whereas the batch policy must make
all decisions simultaneously and cannot benefit from replanning. This
adaptivity gap unavoidable difference in performance is called the adaptivity gap in the
analysis of algorithms.18
Unfortunately, working with the larger action space inherent to batch
optimization requires significant computational effort. First, computing

254
 11.3. synchronous batch observations

input: dataset D, batch size 𝑏, acquisition function 𝛼 Algorithm 11.1: Sequential simulation.
D0 ← D I initialize fictitious dataset
for 𝑖 = 1 . . . 𝑏 do
𝑥𝑖 ← arg max𝑥 ∈X 𝛼 (𝑥; D 0) I select the next batch member
𝑦ˆ𝑖 ← simulate-observation(𝑥

0
𝑖, D )
D ← D ∪ (𝑥𝑖 , 𝑦ˆ𝑖 )
0 0 I update fictitious dataset
end for
return x

the expected marginal gain (11.7) is more expensive than in the sequential
setting (5.8), as we must now integrate with respect to the joint distri-
bution over outcomes 𝑝 (y | x, D). Thus even evaluating the acquisition
function at a single point entails more work. Additionally, finding the
optimal decision (11.8) requires optimizing this score over a significantly
larger domain than in the sequential analog, a nontrivial task due to its
potentially complex and multimodal nature – see figure 11.3.
Despite these computational difficulties, synchronous batch Bayesian
optimization has enjoyed significant attention from the research com-
munity. We can identify two recurring research thrusts: deriving general
strategies for extending arbitrary sequential policies to batch policies
and deriving batch extensions of specific sequential policies. We provide
a brief survey below.

Batch construction via sequential simulation

Sequential simulation is an efficient strategy for creating batch policies by
simulating multiple steps of an existing sequential policy. Pseudocode for
this procedure is listed in Algorithm 11.1. Given a sequential acquisition
function 𝛼, we choose the first batch member by maximization: first batch member, 𝑥 1

𝑥 1 ∈ arg max 𝛼 (𝑥; D),

𝑥 ∈X

and commit to this choice. We now augment our dataset with the chosen
point and a fictitious observed value 𝑦ˆ1 , forming D1 = D ∪ (𝑥 1, 𝑦ˆ1 ) , fictitious observation at 𝑥 1 , 𝑦ˆ 1
then maximize the acquisition function again to choose the second point:
second batch member, 𝑥 2
𝑥 2 ∈ arg max 𝛼 (𝑥; D1 ).
𝑥 ∈X
We proceed in this manner until the desired batch size has been reached.
Sequential simulation entails 𝑏 optimization problems on X rather than a
single problem on X 𝑏, which can be a significant computational savings.
When the sequential policy represents one-step lookahead, sequen- greedy approximation of one-step lookahead
tial simulation can be regarded as a natural greedy approximation to the
one-step batch policy via the decomposition in (11.10): whereas the opti-
mal batch policy must maximize this score jointly, sequential simulation
maximizes the score pointwise, fixing each point once chosen.
This procedure requires some mechanism for generating fictitious
observations as we build the batch. ginsbourger et al. described two

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 255

 extensions and related settings

expected improvement, 𝛼 ei

𝑥1

𝑥6 𝑥5 𝑥3 𝑥4 𝑥2 𝑥7

Figure 11.4: Sequential simulation using the expected improvement (7.2) policy and the kriging believer (11.11) imputation
strategy. The first point is selected by maximizing expected improvement, and we condition the model on an
observation equal to the posterior mean (top panel). We then maximize the updated expected improvement,
condition, and repeat as desired. The bottom panel indicates the locations of several further points selected in
this manner.

19 d. ginsbourger et al. (2010). Kriging is well- simple heuristics that have been widely adopted.19 Perhaps the most
suited to parallelize optimization. In: Compu- natural option is to impute the expected value of each observation, a
tational Intelligence in Expensive Optimization
Problems.
heuristic ginsbourger et al. dubbed the kriging believer strategy:

𝑦ˆ = 𝔼[𝑦 | 𝑥, D]. (11.11)

kriging believer heuristic This has the effect of fixing the posterior mean of the objective function
throughout simulation. An even simpler option is to impute a constant
value independent of the chosen point, which the authors called the
constant liar heuristic constant liar strategy:
𝑦ˆ = 𝑐. (11.12)
Although this might seem silly, this has the advantage of being model
independent and has demonstrated surprisingly good performance in
practice. Three natural options for the constant, ranging from the most
optimistic to most pessimistic, are to impute the maximum, mean, or
minimum of the known observed values y.
example and discussion Seven steps of sequential simulation with the expected improvement
acquisition function (7.2) and the kriging believer strategy (11.11) are
demonstrated in figure 11.4. The selected points appear reasonable: the

256
 11.3. synchronous batch observations

expected improvement, 𝛼 ei

𝑥1

penalty, 𝜑 (𝑥; 𝑥 1 )
1

0
𝑥1

𝑥7 𝑥3 𝑥4 𝑥6 𝑥2 𝑥5

Figure 11.5: Batch construction via local penalization of the expected improvement acquisition function (7.2). We select
the first point by maximizing the expected improvement, after which the acquisition function is multiplied by
a penalty factor discouraging future batch members in that area. We then maximize the updated acquisition
function, penalize, and repeat as desired. The bottom panel indicates the locations of several further points
selected in this manner.

first two exploit the best-seen observation (and are near optimal for a
batch size of two; see figure 11.3), the next two exploit another local
optimum, and the remaider explore the domain.

Batch construction via local penalization

gonzález et al. proposed another general mechanism for extending a
given sequential policy (defined by the acquisition function 𝛼) to the 20 j. gonzález et al. (2016a). Batch Bayesian Opti-
batch setting.20 Like sequential simulation, we select the first point by mization via Local Penalization. aistats 2016.
maximizing the acquisition function:
𝑥 1 ∈ arg max 𝛼 (𝑥; D).
𝑥 ∈X

We then incorporate a multiplicative penalty 𝜑 (𝑥; 𝑥 1 ) into the acquisition penalty from selecting 𝑥 1 , 𝜑 (𝑥; 𝑥 1 )
function discouraging future batch members from being in a neighbor-
hood of the initial point. This penalty is designed to avoid redundancy
between batch members without disrupting the differentiability of the
original acquisition function. gonzález et al. describe one simple and
effective penalty function from an estimate of the global maximum and

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 257

 extensions and related settings

21 Any continually differentiable function on a Lipschitz constant of the objective function.21 We now select the second
compact set is Lipschitz continuous. batch member by maximizing the penalized acquisition function

𝑥 2 ∈ arg max 𝛼 (𝑥; D) 𝜑 (𝑥; 𝑥 1 ),

𝑥 ∈X

after which we apply another penalty and continue in this manner

as desired. This process is illustrated in figure 11.5. Comparing with
sequential simulation, the first batch members are very similar; however,
there is some divergence in the final stages, with local penalization
preferring to revisit the local optimum on the right.
Like sequential simulation, local penalization also entails 𝑏 optimiza-
tion problems on X and is in fact even faster than sequential simulation,
as the objective function model does not need to be updated along the
way.

Approximation via Monte Carlo integration

approximate computation for sequential If we wish to proceed via joint optimization of (11.7) rather than one
one-step lookahead: § 8.5, p. 171 of the above heuristics, we will often face the complication that the
expectation with respect to the observed values y is intractable. We can
offer some advice for approximating this quantity and its gradient for
22 For more discussion on this method, see a Gaussian process model coupled with an exact or additive Gaussian
noise observation model. This abbreviated discussion will closely follow
j. t. wilson et al. (2018). Maximizing acqui-
sition functions for Bayesian optimization. the approach for the sequential case.22
neurips 2018. First we write the one-step marginal gain (11.7) as an expectation of
the marginal gain in utility Δ(x, y) = 𝑢 (D 0) − 𝑢 (D) with respect to a
multivariate normal belief on the observations (2.20):
∫
𝛽 (x; D) = Δ(x, y) N (y; 𝝁, S) dy. (11.13)

23 Gauss–Hermite quadrature, as we recom- We may approximate this expectation via Monte Carlo integration23 by
mended in the one-dimensional case, does not sampling from this belief. It is convenient to do so by sampling 𝑛 vectors
scale well with dimension.
{z𝑖 } from a standard normal distribution, then transforming these via

z𝑖 ↦→ 𝝁 + 𝚲z𝑖 = y𝑖 , (11.14)

where 𝚲 is the Cholesky factor of S: 𝚲𝚲> = S. Monte Carlo estimation

now gives
1 ∑︁
𝑛
𝛽 (x; D) ≈ Δ(x, y𝑖 ). (11.15)
𝑛 𝑖=1

Monte Carlo approximation of gradient As in the sequential case, we may reuse these samples to approximate
the gradient of the acquisition function with respect to the proposed
observation locations, under mild assumptions (c.4–c.5):
   
1 ∑︁ 𝜕Δ
𝑛
𝜕𝛽 𝜕Δ 𝜕Δ 𝜕Δ 𝜕𝝁 𝜕𝚲
≈ (x, y𝑖 ); (x, y𝑖 ) = + + z𝑖 ,
𝜕x 𝑛 𝑖=1 𝜕x 𝜕𝑥 𝑗 𝜕𝑥 𝑗 y 𝜕y 𝜕𝑥 𝑗 𝜕𝑥 𝑗
(11.16)

258
 11.3. synchronous batch observations

where, in the second expression, we account for the dependence of

the y samples on x through the transformation (11.14). The gradient 24 i. murray (2016). Differentiation of the
of the Cholesky factor 𝚲 can be computed efficiently via automatic Cholesky decomposition. arXiv: 1602.07527
[stat.CO].
differentiation.24
We will spend the remainder of this section discussing the details of
explicit batch extensions of popular sequential policies.

Expected improvement and knowledge gradient

Both the expected improvement and knowledge gradient policies have expected improvement: § 7.3 p. 127
been extended to the batch case, closely following the decision-theoretic knowledge gradient: § 7.4 p. 129
approach outlined above.
Unlike its sequential counterpart, computation of batch expected
improvement for Gaussian process models is rather involved, even in the
case of exact observation. The primary challenge is evaluating the ex-
pectation of a multivariate truncated normal distribution, a computation
whose difficulty increases rapidly with the batch size. There are exact
formulas to compute batch expected improvement25 and its gradient26 25 c. chevalier and d. ginsbourger (2013). Fast
based on the moment-generating function for the truncated multivariate Computation of the Multi-points Expected Im-
provement with Applications in Batch Selec-
normal derived by tallis;27 however, these formulas require 𝑏 evalua- tion. lion 7.
tions of the 𝑏-dimensional and 𝑏 2 evaluations of the (𝑏 − 1)-dimensional
26 s. marmin et al. (2015). Differentiating the Mul-
multivariate normal cdf, itself a notoriously difficult computation that tipoint Expected Improvement for Optimal
can only effectively be approximated via Monte Carlo methods in dimen- Batch Design. mod 2015.
sion greater than four. This limits the utility of the direct approach to 27 g. m. tallis (1961). The Moment Generating
relatively small batch sizes, perhaps 𝑏 ≤ 10. Function of the Truncated Multi-normal Dis-
For larger batch sizes, ginsbourger et al. proposed sequential simu- tribution. Journal of the Royal Statistical Soci-
ety Series B (Methodological) 23(1):223–229.
lation, and found the constant liar strategy (11.12) using the “optimistic”
estimate 𝑦ˆ = max y to deliver good empirical performance in simula-
tion.28 wang et al. proposed an efficient alternative: joint optimization 28 d. ginsbourger et al. (2010). Kriging is well-
of batch expected improvement via multistart stochastic gradient as- suited to parallelize optimization. In: Compu-
tational Intelligence in Expensive Optimization
cent using the Monte Carlo estimators in (11.15–11.16).29 The authors Problems.
demonstrated this procedure scaling up to batch sizes of 𝑏 = 128 with 29 j. wang et al. (2020a). Parallel Bayesian Global
impressive performance and runtime compared with exact computation Optimization of Expensive Functions. Opera-
and sequential simulation. tions Research 68(6):1850–1865.
A similar approach for approximating the batch knowledge gradi-
ent policy was described by wu and frazier.30 The overall approach 30 j. wu and p. i. frazier (2016). The Parallel
is effectively the same: multistart stochastic gradient ascent relying on Knowledge Gradient Method for Batch Bayes-
ian Optimization. neurips 2016.
(11.15–11.16). An additional complication in estimating the batch knowl-
edge gradient is the global optimization inherent to the global reward
(6.5). wu and frazier suggest a discretization approach where the global
reward is estimated on a dynamically managed discrete set of points
drawn via Thompson sampling from the objective function posterior. Thompson sampling: 7.9, p. 148
The batch expected improvement acquisition function is illustrated
for an example scenario in figure 11.3, and batch knowledge gradient for
the same scenario in figure 11.6. In both cases, the optimal batch measures
on either side of the previously best-seen point; however, knowledge
gradient is more tolerant of batches containing only one observation in
this region.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 259

 extensions and related settings

observations
posterior mean
posterior 95% credible interval

batch knowledge gradient

optimal batch
𝑥0
𝑥

𝑥
𝑥0

Figure 11.6: The batch knowledge gradient acquisition function for an example scenario. The optimal batch exploits the
local optimum, but any batch containing at least one point in that neighborhood is near-optimal.

Mutual information with 𝑥 ∗

mutual information with 𝑥 ∗ : § 7.6, p. 139 Compared with the difficulty faced in computing (or even approximating)
batch analogs of the expected improvement and knowledge gradient
predictive entropy search: § 8.8, p. 180 acquisition functions, extending the predictive entropy search acquisition
function to the batch setting is relatively straightforward.31 The mutual
31 a. shah and z. ghahramani (2015). Parallel information between the observed values y and the location of the global
Predictive Entropy Search for Batch Global Op- maximum 𝑥 ∗ is (compare with (8.36)):
timization of Expensive Objective Functions.
neurips 2015.  
𝛽𝑥 ∗ (x; D) = 𝐻 [y | x, D] − 𝔼 𝐻 [y | x, 𝑥 ,∗ D] | x, D . (11.17)

The first term is the differential entropy of a multivariate normal and

may be computed in closed form (a.16). The second term is somewhat
difficult to approximate, but no innovation is required in the batch setting
beyond the machinery already developed for the sequential case. We may
Thompson sampling: 7.9, p. 148 approximate the expectation with respect to 𝑥 ∗ via Thompson sampling
and may approximate 𝑝 (y | x, 𝑥 ,∗ D) as a multivariate normal following
the expectation propagation approach described previously.

Probability of improvement
probability of improvement: § 7.5, p. 131 Batch probability of improvement has received relatively little attention,
32 See § 7.5, p. 134.
but jones proposed one simple option in the context of threshold se-
lection.32, 33 The idea is to find the optimal sequential decisions using a
33 d. r. jones (2001). A Taxonomy of Global Op-
timization Methods Based on Response Sur-
range of improvement thresholds, representing a spectrum exploration–
faces. Journal of Global Optimization 21(4):345– exploitation tradeoffs. jones then recommends a simple clustering pro-
383. cedure to remove redundant points, resulting in a batch (of variable size)

260
 11.3. synchronous batch observations

reflecting diversity in location and behavior. A compelling aspect of this

procedure is that it is naturally nonmyopic, as each batch is explicitly
constructed to address both immediate and long-term gain.
This approach is illustrated in figure 7.12;34 depending on the aggres- 34 See p. 136
siveness the pruning procedure, the constructed batch would contain
2–4 points chosen from the visible clusters.

Upper confidence bound

Due to the equivalence between the probability of improvement and maximizing an upper confidence bound: § 7.8,
upper confidence bound policies for Gaussian processes (8.22, 8.26), the p. 145
procedure proposed by jones described above may also be used to realize
a simple batch upper confidence bound policy for that model class. In this
case, we would design each batch by maximizing the upper confidence
bound for a range of confidence parameters, clustering, and pruning.
Several more direct batch upper confidence bound policies have 35 t. desautels et al. (2014). Parallelizing
Exploration–Exploitation Tradeoffs in Gaus-
been developed, all variations on a theme. desautels et al. proposed a sian Process Bandit Optimization. Journal of
strategy – dubbed simply batch upper confidence bound (bucb) – based Machine Learning Research 15(119):4053–4103.
on sequential simulation with the kriging believer strategy (11.11).35 Batch 36 n. de freitas et al. (2012b). Exponential Re-
diversity is automatically encouraged: each point added to the batch gret Bounds for Gaussian Process Bandits with
globally reduces the upper confidence bound, most dramatically at the Deterministic Observations. icml 2012.
locations with the most strongly correlated function values.
The bucb algorithm was later refined by several authors to encourage
more exploration, which can improve both empirical and theoretical
performance. Like bucb, we seed each batch with the maximum of the
upper confidence bound (8.25). We now identify the so-called “relevant
region” of the domain, defined to be the set of locations whose upper
confidence bound exceeds the global maximum of the lower confidence The (disconnected) relevant region (darker
blue) for an example Gaussian process. Points
bound.36 The intuition behind this region is that the objective value outside the region (lighter blue) are unlikely
at any point in its complement is – with high probability – lower than to maximize the objective function.
at the point maximizing the lower confidence bound and can thus be
discarded with some confidence; see the illustration in the margin. 37 a. kulesza and b. taskar (2012). Determi-
With the relevant region in hand, we design the remaining batch nantal Point Processes for Machine Learning.
members to promote maximal information gain about the objective Foundations and Trends in Machine Learning
5(2–3):123–286.
on this region. For a Gaussian process, this is intimately related to a
diversity-encouraging distribution known as a 𝑘-determinantal point 38 e. contal et al. (2013). Parallel Gaussian Pro-
cess Optimization with Upper Confidence
process (𝑘-dpp) built from the posterior covariance function.37 We may
Bound and Pure Exploration. ecml pkdd 2013.
proceed by either a simple greedy procedure38 or via more nuanced
39 t. kathuria et al. (2016). Batched Gaussian
maximization or sampling using methods developed for 𝑘-dpps.39
Process Bandit Optimization via Determinan-
These schemes are all backed by strong theoretical analysis, including tal Point Processes. neurips 2016.
sublinear cumulative regret (10.3) bounds under suitable conditions.

Thompson sampling
The stochastic nature of Thompson sampling enables trivial batch con-
struction by drawing 𝑏 independent samples of the location of the global
maximum (7.19):
{𝑥𝑖 } ∼ 𝑝 (𝑥 ∗ | D).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 261

 extensions and related settings

40 j. m. hernández-lobato et al. (2017). Paral- A remarkable advantage of this policy is that the samples may be gener-
lel and Distributed Thompson Sampling for ated entirely in parallel, which allows linear scaling to arbitrarily large
Large-scale Accelerated Exploration of Chem-
ical Space. icml 2017.
batch sizes. Batch Thompson sampling has delivered impressive perfor-
mance in a real-world setting with batch sizes up to 𝑏 = 500,40 and is
41 k. kandasamy et al. (2018). Parallelised Bayes-
ian Optimisation via Thompson Sampling. ais- backed by theoretical guarantees on the asymptotic reduction of the
tats 2018. simple regret (10.1).41

11.4 asynchronous observation with pending experiments

Some situations allow parallel observation with asynchronous execution.
For example, when optimizing the result of a computational simula-
tion, access to more than one cpu core (or even better, a cluster of
machines) could enable many simulations to be run in parallel. To max-
imize throughput, we could immediately start a new simulation upon
the termination of a previous job, without waiting for the other running
processes to finish. An effective optimization policy for this setting must
account for the pending experiments when designing each observation.
We may consider a general case where we wish to design a batch
of experiments x ∈ X 𝑏 when another batch of experiments x0 ∈ X 𝑏 is
0

under current evaluation, where the number of running and pending

42 Perhaps the most prevalent case in practice experiments may be arbitrary.42 Here the action space for the current
will be 𝑏 = 1 assuming job termination is gov- decision is A ∈ X 𝑏, and we must make the decision under uncertainty
erned by a simple random process.
both in the observations resulting from the chosen batch, y, and the
observations resulting from the pending experiments, y.0
The one-step expected gain in utility from a set of proposed experi-
ments and the pending experiments is
 
𝛽 1 (x; x,0 D) = 𝔼 𝑢 (D1 ) | x, x,0 D − 𝑢 (D),

where D1 = D ∪ (x, y, x,0 y0) . This entails an expectation with respect
to the unknown values y and y0:
∬
𝛽 1 (x; D) + 𝑢 (D) = 𝑢 (D1 ) 𝑝 (y, y0 | x, x,0 D) dy dy.0

reduction to synchronous case This is simply the one-step marginal gain for the combined batch x ∪ x0
from the synchronous case (11.7)! The only difference with respect to
one-step lookahead is that we can only maximize this score with respect
to x as we are already committed to the pending experiments. Thus a
one-step lookahead policy for the asynchronous case can be reduced to
maximizing the corresponding score from the synchronous case with
some batch members fixed. This reduction has been pointed out by
numerous authors, and effectively every batch policy discussed above
may be modified with little effort to work in the asynchronous case.
Moving beyond one-step lookahead may be extremely challenging,
however, due to the implications of uncertainty in the order of termina-
tion for pending experiments. A full treatment would require a model for
the time of termination and accounting for how the decision tree may
branch after the present decision. Exact computation of even two-step

262
 11.5. multifidelity optimization

lookahead is likely intractable in most practical situations, but rollout

might offer one path forward. rollout: § 5.3, p. 102

11.5 multifidelity optimization

In Bayesian optimization, we typically assume that observations of the
objective function are expensive and should be made as sparingly as
possible. However, some scenarios offer a potential shortcut: indirect
inspection of the system of interest via a cheaper surrogate, such as the
output of a computer simulation. In some cases, we may even have access
to multiple surrogates of varying cost and fidelity. It is tempting to try to
accelerate optimization using these surrogates to guide the search. This
is the inspiration for multifidelity optimization, a cost-aware extension
In multifidelity optimization, we wish to
of optimization that has received significant attention. optimize an expensive objective function
As a motivating example, consider a problem from materials science (blue) aided by access to cheaper – but still
where we wish to optimize the properties of a material as a function of informative – surrogates (red).
its composition, process parameters, etc. Materials scientists have sev-
eral mechanisms available43 to investigate a proposed material, ranging materials science applications: appendix d,
from relativelyinexpensive computer simulations (molecular dynamics, p.313
density functional theory, etc.) to extremely expensive synthesis and
characterization in a laboratory. State-of-the-art materials discovery 43 “Relatively” should be stressed; these simula-
pipelines rely on these computational surrogates to winnow the search tions can be quite expensive in absolute terms,
but still much cheaper than synthesis.
space for experimental campaigns.
Automated machine learning provides another motivating example.
Suppose we wish to tune the hyperparameters of a model by minimizing
validation error after training on a large training set. Although training 44 a. klein et al. (2015). Towards efficient Bayes-
on the full dataset may be costly, we can estimate performance by training ian Optimization for Big Data. Bayesian Opti-
mization Workshop, neurips 2015.
on only a subset of the data44 or by terminating the training procedure
early.45 We may reasonably hope to accelerate hyperparameter tuning 45 k. swersky et al. (2014). Freeze–Thaw
by exploiting these noisy, but cost-effective surrogates. Bayesian Optimization. arXiv: 1406 . 3896
[stat.ML].
We will outline a decision-theoretic approach to multifidelity opti-
mization below. The complexity of this setting will require readdressing
every step of the procedure outlined at top of this chapter. We will focus
on the first three steps, as deriving the optimal policy is mechanical once
the model and decision problem are completely specified.

Formalization of problem and action space

Suppose that in addition to the objective function 𝑓, we have access to
one-or-more surrogate functions {𝑓𝑖 } : X → ℝ, indexed by a parameter surrogate functions, {𝑓𝑖 }
𝑖 ∈ I. Most often we take the surrogate functions to form a discrete set, surrogate index set, I
but in some cases we may wish to consider multidimensional and/or
continuous surrogate spaces.46 We denote the objective function itself objective function, 𝑓∗
with the special index ∗ ∈ I, writing 𝑓 = 𝑓∗ . We consider an optimization
scenario where we may design each observation to be either of the 46 k. kandasamy et al. (2017). Multi-fidelity
objective or a surrogate as we see fit, by selecting a location 𝑥 ∈ X for Bayesian Optimisation with Continuous Ap-
proximations. icml 2017.
our next observation and an index 𝑖 ∈ I specifying the desired surrogate.
The action space for each such decision is A = X × I. action space, A = X × I

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 263

 extensions and related settings

Modeling surrogate functions and observations

If surrogate observations are to be useful, they must provide information
about the objective function, and the relationship between the objective
and its surrogates is captured by a joint

model over their values. We first


joint prior process, 𝑝 {𝑓𝑖 } design a joint prior process 𝑝 {𝑓𝑖 } specifying the expected structure
of each individual function and the nature of correlations between the
functions. Next we must create an observation model linking the value
𝑦 observed at a point [𝑥, 𝑖] to the underlying function value 𝜙 = 𝑓𝑖 (𝑥):
observation model, 𝑝 (𝑦 | 𝑥, 𝑖, 𝜙) 𝑝 (𝑦 | 𝑥, 𝑖, 𝜙). Now, given a set of observed data

D, we may derive the pos-


posterior distribution, 𝑝 {𝑓𝑖 } | D terior belief over the functions, 𝑝 {𝑓𝑖 } | D , and the posterior predictive
posterior predictive distribution, distribution, 𝑝 (𝑦 | 𝑥, 𝑖, D), with which we can reason about proposed
𝑝 (𝑦 | 𝑥, 𝑖, D) observations.
joint Gaussian processes: § 2.4, p. 26 In practice, the joint prior process is usually a joint Gaussian process
over the objective and its surrogates. The primary challenge in crafting
a joint Gaussian process is in defining cross-covariance functions

𝐾𝑖 𝑗 = cov[𝑓𝑖 , 𝑓𝑗 ]

that adequately encode the correlations between the functions of inter-

47 m. a. álvarez et al. (2012). Kernels for Vector- est, which can take some care to ensure the resulting joint covariance
Valued Functions: A Review. Foundations and
function over the collection {𝑓𝑖 } is positive definite.
Trends in Machine Learning 4(3):195–266.
Fortunately, a great deal of effort has been invested in developing
48 k. ulrich et al. (2015). gp Kernels for Cross- this model class into a flexible and expressive family.47, 48 One simple
Spectrum Analysis. neurips 2015.
construction offering some intuition is the separable covariance


𝐾 [𝑥, 𝑖], [𝑥 ,0 𝑖 0] = 𝐾X (𝑥, 𝑥 0) 𝐾I (𝑖, 𝑖 0), (11.18)

which decomposes the joint covariance into a covariance function on the

shared domain covariance, 𝐾X domain 𝐾X shared by each individual function and a covariance function
cross-function covariance, 𝐾I between the functions, 𝐾I (which would be a covariance matrix if I is
finite). In this construction, the marginal covariance and cross-covariance
functions are all scaled versions of 𝐾X , with the 𝐾I covariance scaling
each marginal belief and encoding (constant) cross-correlations across
49 See pp. 28–29. functions as well. Figures 2.5 and 2.649 illustrate the behavior of this
model for two highly correlated functions on a shared one-dimensional
domain.

Defining a utility function

We have now addressed steps 2 and 3 of our general procedure for mul-
tifidelity optimization, by identifying the expanded action space implied
by the problem and determining how to reason about the potential out-
comes of these actions. Before we can proceed with deriving a policy,
however, we must establish preferences over outcomes with a utility
function 𝑢 (D). It is difficult to provide specific guidance for this choice,
as these preferences are inherently bound to a given situation. One natu-
ral approach would be to choose a cost-aware utility function measuring
optimization progress limited to the objective function alone, adjusted

264
 11.5. multifidelity optimization

objective, 𝑓∗ surrogate, 𝑓1

cost-adjusted expected improvement next observation location

Figure 11.7: Cost-adjusted multifidelity expected improvement for a toy scenario. The objective (left) and its surrogate
(right) are modeled as a joint Gaussian process with marginals equivalent to the example from figure 7.1 and
constant cross-correlation 0.8. The cost of observation was assumed to be ten times greater for the objective
than its surrogate. Maximizing the cost-adjusted multifidelity expected improvement elects to continue
evaluating the surrogate.

for the variable costs of each observation obtained. For example we

might quantify the cost of observing the objective function or each of
the surrogate functions with values {𝑐𝑖 }, and adjust a data utility 𝑢 0(D) observation costs, {𝑐𝑖 }
appropriately, by defining
∑︁
𝑢 (D) = 𝑢 0(D) − 𝑐𝑖 .
(𝑥,𝑖) ∈D

With an appropriate utility function in hand, we can then proceed to

derive the optimal policy as usual.
As an example, we can realize an analog of expected improvement multi-fidelity expected improvement
(7.2) through a suitable redefinition of the simple reward (6.3). Suppose
that at termination we wish to recommend a location visited during op-
timization, with any fidelity,

using a risk-neutral utility. Given a multifi-
delity dataset D = [x, i], y , the expected utility of this recommendation
would be

𝑢 0(D) = max0
𝜇D [𝑥 ,0 ∗] ,
𝑥 ∈x
the maximum of the posterior mean for the objective function at the
observed locations.
Figure 11.7 illustrates one-step lookahead policy with this utility example and discussion
function for a one-dimensional objective function 𝑓∗ (left) and a surro-
gate 𝑓1 (right). The marginal belief about each function is a Gaussian
process identical to our running example from chapter 7; see figure
7.1. These are coupled together via the separable covariance (11.18) with
𝐾I (∗, ∗) = 𝐾I (1, 1) = 1 and cross-correlation 𝐾I (1, ∗) = 0.8. We begin
with three surrogate observations and compute the cost-adjusted ex-
pected improvement as described above, where the cost of observing
the objective was set to ten times that of the surrogate. In this case, the
one-step optimal decision is to continue evaluating the surrogate around
the best-seen surrogate observation.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 265

 extensions and related settings

objective, 𝑓∗

Figure 11.8: A simulation of optimization with the cost-adjusted multifidelity expected improvement starting from the
scenario in figure 11.7. We simulate sequential observations of either the objective or the surrogate, illustrated
using the running tick marks. The light gray marks correspond to surrogate observations and the heavy
black marks objective observations. The optimum was found after 10 objective and 32 surrogate observations,
marked in heavy red. The prior and posterior of the objective function conditioned on all observations are
also shown.

50 d. huang et al. (2006a). Sequential kriging op- Figure 11.8 simulates sequential multifidelity optimization using this
timization using multiple-fidelity evaluations. policy; here the optimum was discovered after only 10 evaluations of
Structural and Multidisciplinary Optimization the objective, guided by 32 observations of the cheaper surrogate. A
32(5):369–382.
remarkable feature we can see in the posterior is that all evaluations
51 v. picheny et al. (2013a). Quantile-Based Op- made of the objective function are above the prior mean, nearly all with
timization of Noisy Computer Experiments
With Tunable Precision. Technometrics 55(1):
𝑧-scores of approximately 𝑧 = 1 or greater. This can be ascribed not
2–13. to extreme luck, but rather to efficient use of the surrogate to rule out
52 j. wu et al. (2019). Practical Multi-fidelity
regions unlikely to contain the optimum.
Bayesian Optimization for Hyperparameter Multifidelity Bayesian optimization has enjoyed sustained interest
Tuning. uai 2019. from the research community, and numerous policies have been available.
53 k. kandasamy et al. (2016). Gaussian Process These include adaptations of the expected improvement,50, 51 knowledge
Bandit Optimisation with Multi-fidelity Eval- gradient,52 upper confidence bound,53, 54 and mutual information with
uations. neurips 2016. 𝑥 ∗ 55, 56 and 𝑓 ∗ 57 acquisition functions, as well as novel approaches.58
54 k. kandasamy et al. (2017). Multi-fidelity
Bayesian Optimisation with Continuous Ap-
proximations. icml 2017. 11.6 multitask optimization
55 k. swersky et al. (2013). Multi-Task Bayesian
Optimization. neurips 2013. Multitask optimization addresses the sequential or simultaneous opti-
mization of multiple objectives {𝑓𝑖 } : X → ℝ representing performance

266
 11.6. multitask optimization

on related tasks. Like multifidelity optimization, the underlying idea

56 y. zhang et al. (2017). Information-Based
in multitask optimization is that if performance on the various tasks is Multi-Fidelity Bayesian Optimization. Bayes-
correlated as a function of the input, we may accelerate optimization by ian Optimization Workshop, neurips 2017.
transferring information between tasks. 57 s. takeno et al. (2020). Multi-fidelity Bayesian
As a motivating example of sequential multitask optimization, con- Optimization with Max-value Entropy Search
sider a web service wishing to retune the parameters of an ad placement and its Parallelization. icml 2020.
algorithm on a regular basis to maximize revenue in the current climate. 58 j. song et al. (2019). A General Framework
for Multi-fidelity Bayesian Optimization with
Here the revenue at each epoch represents the different tasks to be op-
Gaussian Processes. aistats 2019.
timized, which are optimized individually one after another. Although
we could treat each optimization problem separately, they are clearly
related, and with some care we may be able to use past performance to
provide a “warm start” to each new optimization problem rather than
start from scratch.
We may also consider the simultaneous optimization of performance simultaneous tasks
across tasks. For example, a machine learning practitioner may wish to
tune model hyperparameters to maximize the average predictive perfor-
mance on several related datasets.59 A naïve approach would formulate 59 This could also be formulated as a scalarized
the problem as maximizing a single objective defined to be the average version of a multiobjective optimization prob-
lem, discussed in the next section.
performance across tasks, with each evaluation entailing retraining the
model for each dataset. However, this would be potentially wasteful, as
we may be able to eliminate poorly performing hyperparameters with
high confidence after training on a fraction of the datasets. A multi-
task approach would model each objective function separately (perhaps
jointly) and consider evaluations of single-task performance to efficiently
maximize the combined objective.
swersky et al. described a particularly clever realization of this idea:
selecting model hyperparameters via cross validation.60 Here we recog- 60 k. swersky et al. (2013). Multi-Task Bayesian
nize the predictive performance on each validation fold as being corre- Optimization. neurips 2013.
lated due to shared training data across folds. If we can successfully share
information across folds, we may potentially accelerate cross validation
by iteratively selecting (hyperparameter, fold index) pairs rather than
training proposed hyperparameters on every fold each time.

Formulation and approach

Let {𝑓𝑖 } : X → ℝ be the set of objective functions we wish to consider, modeling task objectives and observations
representing performance on the relevant tasks. As with multifidelity
optimization,

the key enabler of multitask optimization is a joint model
𝑝 {𝑓𝑖 } over the tasks and a joint observation model 𝑝 (𝑦 | 𝑥, 𝑖, 𝜙) over
evaluations thereof; this joint model allows us to share information joint gps for modeling multiple functions:
between the tasks. This could, for example, take the form of a joint § 11.5, p. 264
Gaussian process, as discussed previously.
Once this model has been chosen, we can turn to the problem of sequential tasks
designing a multitask optimization policy. If each task is to be solved
one at a time, a natural approach would be to design the utility function
to ultimately evaluate performance on that task only. In this case, the
problem devolves to single-objective optimization, and we may use any
of the approaches discussed earlier in the book to derive a policy. The

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 267

extensions and related settings

current task, 𝑓

Figure 11.9: A demonstration of sequential multitask optimization. Left: a prior distribution over an objective function
along with 13 observations selected by maximizing expected improvement, revealing the global maximum with
the last evaluation. Right: the posterior distribution over the same objective conditioned on the observations
of the two functions (now interpreted as related tasks) in figure 11.8. The global maximum is now found after
three observations due to the better informed prior.

only difference is that the objective function model is now informed from
our past experience with other tasks; as a result, our initial optimization
decisions can be more targeted.
This procedure is illustrated in figure 11.9. Both panels illustrate
the optimization of an objective function by sequentially maximizing
expected improvement. The left panel begins with no information and
locates the global optimum after 13 evaluations. The right panel begins
the process instead with a posterior belief about the objective conditioned
on the data obtained from the two functions in figure 11.8, modeled as
related tasks with cross-correlation 0.8. Due to the better informed initial
belief, we now find the global optimum after only three evaluations.
simultaneous tasks The case of simultaneous multitask optimization – where we may
evaluate any task objective with each observation – requires somewhat
more care. We must now design a utility function capturing our joint
performance across the tasks and design each observation with respect
to this utility. One simple option would be to select utility functions
{𝑢𝑖 } quantifying performance on each task separately and then take a
weighted average: ∑︁
𝑢 (D) = 𝑤𝑖 𝑢𝑖 (D).
𝑖
This could be further adjusted for (perhaps task- and/or input-dependent)
observation costs if needed. Now we may write the expected marginal
gain in this combined utility as a weighted average of the expected
marginal gain on each separate task. One-step lookahead would then
maximize the weighted acquisition function
∑︁

𝛼 [𝑥, 𝑖]; D) = 𝑤𝑖 𝛼𝑖 [𝑥, 𝑖]; D
𝑖

over all possible observation location–task pairs [𝑥, 𝑖], where 𝛼𝑖 is the
one-step expected marginal gain in 𝑢𝑖 . Note that observing a single task

268
 11.7. multiobjective optimization

𝑓1 𝑓2

𝑥1 𝑥2 𝑥4 𝑥3

Figure 11.10: A simple multiobjective optimization example with two objectives {𝑓1, 𝑓2 } on a one-dimensional domain. We
compare four identified points: 𝑥 1 , the global optimum of 𝑓1 , 𝑥 2 , the global optimum of 𝑓2 , 𝑥 3 , a compromise
with relatively high values of both objectives, and 𝑥 4 , a point with relatively low values of both objectives.

could in fact yield improvement in all utilities due to information sharing

through the joint belief on task objectives.

11.7 multiobjective optimization

Like multitask optimization, multiobjective optimization addresses the
simultaneous optimization of multiple objectives {𝑓𝑖 } : X → ℝ. However,
whereas in multitask optimization we usually seek to identify the global
optimum of each function separately, in multiobjective optimization we
seek to identify points jointly optimizing all of the objectives. Of course,
this is not possible unequivocally unless all of the maxima happen to
coincide, as we may need to sacrifice the value of one objective in order
to increase another. Instead, we may consider the optimization of various
tradeoffs between the objectives and rely on subjective preferences to
determine which option is preferred in a given scenario. Multiobjective
optimization may then be posed as the identification of one or more
optimal tradeoffs among the objectives to support this analysis.
A classic example of multiobjective optimization can be found in
finance, where we seek investment portfolios optimizing tradeoffs be- risk: standard deviation of return
tween risk (often captured by the standard deviation of return) and
reward (often captured by the expected return). Generally, investments reward: expected value of return
with higher risk yield higher reward, but the optimal investment strategy
depends on the investor’s risk tolerance – for example, when capital 61 In modern portfolio theory the term “efficient
preservation is paramount, low-risk, low-reward investments are pru- frontier” is more common for the equivalent
concept.
dent. The set of investment portfolios maximizing reward for any given
risk is known as the Pareto frontier,61 which jointly span all rational Pareto frontier
solutions to the given problem. We generalize this concept below.

Pareto optimality
To illustrate the tradeoffs we may need to consider during multiobjec-
tive optimization, consider the two objectives in figure 11.10. The first
objective 𝑓1 has its global maximum at 𝑥 1 , which nearly coincides with
the global minimum of the second objective 𝑓2 . The reverse is true in
the other direction: the global maximum of the second objective, 𝑥 2 ,

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 269

 extensions and related settings

𝑓1 𝑓2

𝑥1 𝑥2 𝑥3

Figure 11.11: The objectives from figure 11.10 with the Pareto optimal solutions highlighted. All points along the intervals
marked on the horizontal axis are Pareto optimal, with the highlighted corresponding objective values forming
the Pareto frontier (see margin).

achieves a relatively low value on the first objective. Neither point can
𝑓2 be preferred over the other on any objective grounds, but a rational agent
𝑥2 may have subjective preferences over these two locations depending on
𝑥3
the relative importance of the two objectives. Some agents might even
prefer a compromise location such as 𝑥 3 to either of these points, as it
achieves relatively high – but suboptimal – values for both objectives.
It is clearly impossible to identify an unambiguously optimal location,
even in this relatively simple example. We can, however, eliminate some
locations as plainly subpar. For example, consider the point 𝑥 4 in figure
𝑥1 11.10. Assuming preferences are nondecreasing with each objective, no
rational agent would prefer 𝑥 4 to 𝑥 3 as the latter point achieves higher
𝑥4
value for both objectives. We may formalize this intuition by defining a
𝑓1
partial order on potential solutions consistent with this reasoning.
We will say that a point 𝑥 dominates another point 𝑥 ,0 denoted 𝑥 0 ≺ 𝑥,
The regions dominated by points 𝑥 1 and 𝑥 3
in figure 11.10 are highlighted in blue. 𝑥 4 is if no objective value is lower at 𝑥 than at 𝑥 0 and if at least one objective
dominated by both, and 𝑥 2 by neither. value is higher at 𝑥 than at 𝑥 .0 Assuming preferences are consistent with
nondecreasing objective values, no agent could prefer a dominated point
𝑓2 to any point dominating it. This concept is illustrated in the margin for
𝑥2 the example from figure 11.10: all points in the blue regions are dominated,
and in particular 𝑥 4 is dominated by both 𝑥 1 and 𝑥 3 . On the other hand,
𝑥3 none of 𝑥 1 , 𝑥 2 , or 𝑥 3 is dominated by any of the other points.
A point 𝑥 ∈ X that is not dominated by any other point is called
Pareto optimal, and the image of all Pareto optimal points is called the
Pareto frontier. The Pareto frontier is a central concept in multiobjective
optimization – it represents the set of all possible solutions to the problem
consistent with weakly monotone preferences for the objectives. Figure
𝑥1
11.11 shows the Pareto optimal points for our example from figure 11.10,
which span four disconnected intervals. We may visualize the Pareto
𝑓1 frontier by plotting the image of this set, as shown in the margin.
There are several approaches to multiobjective optimization that
The Pareto frontier for the scenario in figures differ in when preferences among competing solutions are elicited.62
11.10–11.11. The four components correspond
So-called a posteriori methods seek to identify the entire Pareto frontier
to the highlighted intervals in figure 11.11.
for a given problem, with preferences among possible solutions to be
62 k. m. miettinen (1998). Nonlinear Multiobjec-
determined afterwards. In contrast, a priori methods assume that pref-
tive Optimization. Kluwer Academic Publish-
ers. erences are already predetermined, allowing us to seek a single Pareto

270
 11.7. multiobjective optimization

optimal solution consistent with those preferences. Bayesian realizations

of both types of approaches have been realized, as we discuss below.

Formalization of decision problem and modeling objectives

Nearly all Bayesian multiobjective optimization procedures model each
decision as choosing a location 𝑥 ∈ X, where we make an observation
of every objective function. We could also consider a setting analogous
to multitask or multifidelity optimization where we observe only one
objective at a time, but this idea has not been sufficiently explored. For the
following discussion, we will write y for the vector-valued observation vector of observations at 𝑥, y
resulting from an observation at a given point 𝑥, with 𝑦𝑖 being associated
with objective 𝑓𝑖 .
As in the previous two sections, we build a joint model for the ob-

jectives {𝑓𝑖 } and our observations of them
via a prior process 𝑝 {𝑓𝑖 }
and an observation model 𝑝 y | 𝑥, {𝜙𝑖 } . The models are usually taken
to be independent Gaussian processes on each objective combined with 63 j. svenson and t. santner (2016). Multiobjec-
standard observation models. Joint Gaussian processes could also be tive optimization of expensive-to-evaluate de-
terministic computer simulator models. Com-
used when appropriate, but a direct comparison of independent versus putational Statistics and Data Analysis 94:250–
dependent models did not demonstrate improvement when modeling 264.
correlations between objectives, perhaps due to an increased burden in
estimating model hyperparameters.63

Expected hypervolume improvement

64 e. zitzler (1999). Evolutionary Algorithms for
The majority of Bayesian multiobjective optimization approaches are a Multiobjective Optimization: Methods and Ap-
posteriori methods, seeking to identify the entire Pareto frontier or a rep- plications. PhD thesis. Eidgenössische Tech-
resentative portion of it. Many algorithms represent one-step lookahead nische Hochschule Zürich. [§ 3.1]
for some utility function evaluating progress on this task.
One popular utility for multiobjective optimization is the volume 𝑓2
under an estimate of the Pareto frontier, also known as the S metric.64
Namely, given observations of the objectives, we may build a natural 𝑥1
statistical lower bound of the Pareto frontier by eliminating the outcomes
𝑥2
dominated by the observations with high confidence. When observations
are exact, we may simply enumerate the dominated regions and take 𝑥3
their union; when observations are corrupted by noise, we may use a
statistical lower bound of the underlying function values instead.65 This
procedure is illustrated in the margin for our running example, where
we have made exact observations of the objectives at three mutually
nondominating locations; see figure 11.12. The upper-right boundary of r
the dominated region is a lower bound of the true Pareto frontier. 𝑓1
To evaluate progress on mapping out the Pareto frontier, we con-
sider the volume of space dominated by the available observations and A lower bound of the Pareto frontier for our
example given the data in figure 11.12.
bounded below by an identified, clearly suboptimal reference point r; see
the blue shaded area in the margin. The reference point is necessary to 65 m. emmerich and b. naujoks (2004). Meta-
model Assisted Multiobjective Optimisation
ensure the dominated volume does not diverge to infinity. Assuming the Strategies and their Application in Airfoil De-
reference point is chosen such that it will definitely be dominated, this sign. In: Adaptive Computing in Design and
utility is always positive and is maximized when the true Pareto fron- Manufacture vi.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 271

 extensions and related settings

posterior mean, 𝑓1 posterior 95% credible interval, 𝑓1

posterior mean, 𝑓2 posterior 95% credible interval, 𝑓2
observations

Figure 11.12: The posterior belief about

our example objectives from
figure 11.10 given observa-
tions at the marked loca-
tions. The beliefs are sep-
arated vertically (by an ar-
bitrary amount) for clarity,
with the belief over the other
function shown in gray for
reference.

𝑥1 𝑥2 𝑥3

expected hypervolume improvement next observation location

Figure 11.13: The expected hypervolume

improvement acquisition
function for the above
example.

66 m. fleischer (2003). The Measure of Pareto tier is revealed by the observed data.66 Therefore it provides a sensible
Optima: Applications to Multi-objective Meta- measure of progress for a posteriori multiobjective optimization.
heuristics. emo 2003.
The one-step marginal gain in this utility is known as expected hy-
67 m. t. m. emmerich et al. (2006). Single- and pervolume improvement (ehvi)67, 68 and serves as a popular acquisition
Multiobjective Evolutionary Optimization As-
sisted by Gaussian Random Field Metamodels. function. This score is shown in figure 11.13 for our example; the optimal
ieee Transactions on Evolutionary Computation decision attempts to refine the central portion of the Pareto frontier, but
10(4):421–439. many alternatives are almost as favorable due to the roughness of the
68 w. ponweiser et al. (2008). Multiobjective current estimate. Computation of ehvi is involved, and its cost grows
Optimization on a Limited Budget of Evalu- considerably with the number of objectives. The primary difficulty is
ations Using Model-Assisted S-Metric Selec-
enumerating and integrating with respect to the lower bound to the
tion. ppsn x.
Pareto front, which can become a complex region in higher dimensions.
69 k. yang et al. (2019b). Multi-Objective Bayes-
ian Global Optimization using expected hy- However, efficient algorithms are available for computing evhi69 and its
pervolume improvement gradient. Swarm and gradient.70
Evolutionary Computation 44:945–956.
70 k. yang et al. (2019a). Efficient computation
of expected hypervolume improvement us-
Information-theoretic a posteriori methods
ing box decomposition algorithms. Journal of
Several popular information-theoretic policies for single-objective opti-
Global Optimization 75(1):3–34.
mization have been adapted to a posteriori multiobjective optimization.
The key idea behind these methods is to approximate the mutual infor-
mation between a joint observation of the objectives and either the set
of Pareto optimal points (in the domain), X ∗, or the Pareto frontier (in
the codomain), F ∗. These approaches operate by maximizing the predic-
tive form of mutual information (8.35) and largely follow the parallel
single-objective cases in their approximation.

272
 11.7. multiobjective optimization

hernández-lobato et al. proposed maximizing the mutual infor-

mation between the observations y realized at a proposed observation
location 𝑥 and the set of Pareto optimal points X ∗ :
 
𝛼 pesmo (𝑥; D) = 𝐻 [y | 𝑥, D] − 𝔼X ∗ 𝐻 [y | 𝑥, D, X ∗ ] | 𝑥, D ,
calling their policy predictive entropy search for multiobjective optimiza-
tion (pesmo).71 As in the single-objective case, for Gaussian process 71 d. hernández-lobato et al. (2016a). Predic-
models with additive Gaussian noise, the first term of this expression tive Entropy Search for Multi-objective Bayes-
ian Optimization. icml 2016.
can be computed exactly as the differential entropy of a multivariate
normal distribution (a.16). However, the second term entails two com-
putational barriers: computing an expectation with respect to X ∗ and
conditioning our objective function belief on this set. The authors pro-
vide approximations for each of these tasks for Gaussian process models
based on Gaussian expectation propagation; these are reminiscent of the
procedure used in predictive entropy search. predictive entropy search: § 8.8, p. 180
belakaria et al. meanwhile proposed maximizing the mutual infor-
mation with the Pareto frontier F ,∗
 
𝛼 mesmo (𝑥; D) = 𝐻 [y | 𝑥, D] − 𝔼F ∗ 𝐻 [y | 𝑥, D, F ∗ ] | 𝑥, D .
The authors dubbed the resulting policy max-value entropy search for
multiobjective optimization (mesmo).72 This policy naturally shares many 72 s. belakaria et al. (2019). Max-value Entropy
features with the pesmo policy. Again the chief difficultly is in addressing Search for Multi-Objective Bayesian Optimiza-
tion. neurips 2019.
the expectation and conditioning with respect to F ∗ in the second term
of the mutual information; however, both of these tasks are rendered
somewhat easier by working in the codomain. To approximate the expec-
tation with respect to F ,∗ the authors propose a Monte Carlo approach
where posterior samples of the objectives are generated efficiently using
a sparse-spectrum approximation, which are fed into an off-the-shelf, sparse spectrum approximation: § 8.7, p. 178
exhaustive multi-objective optimization routine. Conditioning y on a
realization of the Pareto frontier now entails appropriate truncation of
its multivariate normal belief.

A priori methods and scalarization

When preferences regarding tradeoffs among the objective functions can
be sufficiently established prior to optimization, perhaps with input from
a domain expert, we may reduce multiobjective optimization to a single
objective problem by explicitly maximizing the desired criterion. This
is called scalarization and is a prominent a priori method. Reframing in scalarization
terms of a single objective offers the obvious benefit of allowing us to
appeal to the expansive methodology for that purpose we have built up
throughout the course of this book.
Scalarization is an important component of some a posteriori mut- a posteriori optimization via scalarization
liobjective optimization methods as well. The idea is to construct a family
of exhaustive parametric scalarizations of the objectives such that the
solution to any such problem is Pareto optimal, and that by spanning
the parameter range we may reveal the entire Pareto frontier one point
at a time.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 273

 extensions and related settings

Figure 11.14: A series of linear scalarizations of the example objectives from figure 11.10. The “blue end” of the color
spectrum is 𝑓1 , and the “red end” 𝑓2 , consistent with that plot.

vector of objective values at 𝑥, 𝝓 Let 𝝓 denote the vector of objective function values at an arbitrary
scalarization function, 𝑔 location 𝑥, defining 𝜙𝑖 = 𝑓𝑖 (𝑥). A scalarization function 𝑔 : 𝑥 ↦→ 𝑔(𝝓) ∈ ℝ
maps locations in the domain to scalars determined by their objective val-
ues. We may interpret the output as defining preferences over locations
in a natural manner: namely, if 𝑔(𝝓) > 𝑔(𝝓 0), then the outcomes at 𝑥 are
preferred to those at 𝑥 0 in the scalarization. With this interpretation, a
scalarization function allows us to recast multiobjective optimization as
a single-objective problem by maximizing 𝑔 with respect to 𝑥.
A scalarization function can in principle be arbitrary, and a priori
multiobjective optimization can be framed in terms of maximizing any
such function. However, several tunable scalarization functions have
been described in the literature that may be used in a general context.
𝑓2
A straightforward and intuitive example is linear scalarization:
∑︁
𝑔lin (𝑥; w) = 𝑤 𝑖 𝜙𝑖 , (11.19)
𝑖

where each weight 𝑤𝑖 is nonnegative. A range of linear scalarizations

for our running example is shown in figure 11.14, here constructed to
smoothly interpolate between the two objectives. The maximum of a
linear scalarization is guaranteed to lie on the Pareto frontier; however,
not every Pareto optimal point can be recovered in this manner unless the
frontier is strictly concave. That is the case for our example, illustrated
in the marginal figure. If we model each objective with a (perhaps joint)
𝑓1 Gaussian process, then the induced belief about any linear scalarization
is conveniently also a Gaussian process, so no further modeling would
The marked points represent the maxima of be required for the scalarization function itself.
the similarly colored linear scalarizations in
figure 11.14. Another choice that has seen some use in Bayesian optimization is
augmented Chebyshev scalarization, which augments the linear scalar-
ization (11.19) with an additional, nonlinear term:
 
𝑔ac (𝑥; w, 𝜌) = min 𝑤𝑖 (𝜙𝑖 − 𝑟𝑖 ) + 𝜌 𝑔lin (𝑥; w). (11.20)
𝑖
73 j. knowles (2005). Parego: A Hybrid Algo-
rithm With On-Line Landscape Approxima- Here r is a reference point as above and 𝜌 is a small nonnegative constant;
tion for Expensive Multiobjective Optimiza-
tion Problems. ieee Transactions on Evolution-
knowles for example took 𝜌 = 0.05.73 The augmented Chebyshev scalar-
ary Computation 10(1):50–66. ization function has the benefit that all points on the Pareto frontier can
74 d. golovin and q. zhang (2020). Random Hy- be realized by maximizing with respect to some corresponding setting
pervolume Scalarizations for Provable Multi- of the weights, even if the frontier is nonconcave. golovin and zhang
Objective Black Box Optimization. icml 2020. proposed a similar hypervolume scalarization related to the S-metric.74

274
 11.7. multiobjective optimization

posterior mean, 𝑓1 posterior 95% credible interval, 𝑓1

posterior mean, 𝑓2 posterior 95% credible interval, 𝑓2
observations

1 Pr(𝑥 ⊀ x | D)
Figure 11.15: The probability of nondomi-
nance by the available data
for our running example
0 (above).

Several authors have derived Bayesian methods for a posteriori multi-

objective optimization by solving a series of carefully constructed scalar-
ized problems. knowles for example proposed sampling random weights
for the augmented Chebyshev scalarization (11.20) and optimizing the
resulting objective, repeating this process until satisfied.73 paria et al.
proposed a similar approach incorporating a prior distribution over the
parameters of a chosen scalarization function to allow the user to focus
on an identified region of the Pareto frontier if desired.75 The procedure
then proceeds by repeatedly sampling from that distribution and maxi-
mizing the resulting objective. This is effectively the same as knowles’s
approach, but the authors were able to establish theoretical regret bounds 75 b. paria et al. (2019). A Flexible Framework for
for their procedure. These results were improved by golovin and zhang, Multi-Objective Bayesian Optimization using
Random Scalarizations. uai 2019.
who derived (stronger) theoretical guarantees for a procedure based on
repeatedly optimizing their proposed hypervolume scalarization.

Other approaches
zuluaga et al. outlined an intriguing approach to a posteriori multi-
objective optimization76 wherein the problem was recast as an active 76 m. zuluaga et al. (2016). 𝜀-pal: An Active
learning problem.77 Namely, the authors considered the binary classifi- Learning Approach to the Multi-Objective Op-
timization Problem. Journal of Machine Learn-
cation problem of predicting whether a given observation location was ing Research 17(104):1–32.
(approximately) Pareto optimal or not, then designed observations to
77 b. settles (2012). Active Learning. Morgan &
maximize expected performance on this task. Their algorithm is sup- Claypool.
ported by theoretical bounds on performance and performed admirably
78 v. picheny (2015). Multiobjective optimization
compared with knowles’s algorithm discussed above.73 using Gaussian process emulators via stepwise
picheny proposed a spiritually similar approach also based on se- uncertainty reduction. Statistics and Comput-
quentially reducing a measure of uncertainty in the Pareto frontier.78 ing 25(6):1265–1280.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 275

 extensions and related settings

Given a dataset D = (x, y), we consider the probability that a given

point 𝑥 ∈ X is not dominated by any point in the current dataset (that
is, the probability 𝑥 may lie on the Pareto frontier but not yet be dis-
covered), Pr(𝑥 ⊀ x | D). The integral of this score over the domain
can be interpreted as a measure of uncertainty in the Pareto frontier as
determined by the data, and negating this measure provides a plausible
utility function for a posteriori multiobjective optimization:
∫
𝑢 (D) = − Pr(𝑥 ⊀ x) d𝑥 .

This probability is plotted for our running example in figure 11.15; here,
there is a significant probability for many points to be nondominated
by the rather sparse available data, indicating a significant degree of
uncertainty in our understanding of the Pareto frontier. However, as
the Pareto frontier is increasingly well determined by the data, this
probability will vanish globally and the utility above will tend toward
its maximal value of zero. After motivating this score, picheny proceeds
to recommend designing observations via one-step lookahead.

11.8 gradient observations

Bayesian optimization is often described as a “derivative-free” approach
to optimization, but this characterization is misleading. Although it is
true that Bayesian optimization methods do not require the ability to
observe derivatives, it is certainly not the case that we cannot make use
of such observations when available. In fact, it is straightforward to con-
conditioning a gp on derivative observations: dition a Gaussian process on derivative observations, even if corrupted
§ 2.6, p. 32 by noise, and so from a modeling perspective we are already done.
Of course, ideally, our policy should also consider the acquisition
79 This could be, for example, exact observation of derivative information due to its influence on our belief and the
or additive Gaussian noise. Recall that for a utility of collected data, of which they now form a part. To do so is by
gp on 𝑓, the joint distribution of (𝜙, ∇𝜙) is
multivariate normal (2.28), so for these choices
now relatively simple. A fairly general scheme would assume that an
the predictive distribution of (𝑦, g) is jointly observation at 𝑥 yields a pair of measurements (𝑦, g) respectively related
Gaussian and exact inference is tractable. to (𝜙, ∇𝜙) via a joint observation model.79 We can then compute the
one-step expected marginal gain in utility from observing these values:
∬
 
𝛼 1 (𝑥; D) = 𝑢 (D1 ) − 𝑢 (D) 𝑝 (𝑦, g | 𝑥, D) d𝑦 dg, (11.21)

where the updated dataset D1 will reflect the entire observation (𝑥, 𝑦, g).
Induction on the horizon gives the optimal policy as usual.
Figure 11.16 compares derivative-aware and derivative-unaware ver-
sions of the knowledge gradient (7.4) (assuming exact observation) for an
example scenario. The derivative-aware version dominates the derivative-
unaware one, as the acquisition of more information naturally leads to a
greater expected marginal gain in utility. When derivative information
is unavailable, the optimal decision is to evaluate nearby the previously
best-seen point, effectively to estimate the derivative via finite differ-
encing; when derivative information is available, an observation at this

276
 11.9. stochastic and robust optimization

derivative-aware knowledge gradient next observation location

knowledge gradient

Figure 11.16: The knowledge gradient acquisition function for an example scenario reflecting the expected gain in global
reward (6.5) provided exact observations of the objective function and its derivative. The vanilla knowledge
gradient (7.4) based on an observation of the objective alone is shown for reference.

location yields effectively the same expected gain. However, we can fair
even better in expectation by moving a bit farther astray, where the
derivative information is less redundant.
When working with Gaussian process models in high dimension, scaling of Gaussian process inference: § 9.1,
it may not be wise to augment each observation of the objective with p. 201
a full observation of the gradient due to the cubic scaling of inference.
However, the scheme outlined above opens the door to consider the
observation of any measurement related to the gradient. For example,
we might condition on the value of a directional derivative, reducing the
measurement to a scalar regardless of dimension and limiting compu-
tation. Such a scheme was promoted by wu et al., who considered the
acquisition of a single coordinate of the gradient; this could be extended
to non-axis-aligned directional derivatives without major complication.80 80 j. wu et al. (2017). Bayesian Optimization with
We could also consider a “multifidelity” extension where we weigh vari- Gradients. neurips 2017.
ous possible gradient observations (including none at all) in light of their
expected utility and the cost of acquisition/inference.

11.9 stochastic and robust optimization

In some applications, the performance of a given system configuration
depends on stochastic exogenous factors. For example, consider optimiz-
ing the parameters of a mobile robot’s gait to maximize some tradeoff of
stability, efficiency, and speed. These objectives depend not only on the
gait parameters, but also on the nature of the environment – such as the
composition, slope, etc. of the surface to be traversed – which cannot
be controlled by the robot at the time of performance and may vary
over repeated sessions. In this scenario, we may seek to optimize some
measure of performance accounting for uncertainty in the environment.
To model this and related scenarios, we may consider an objective
function 𝑔(𝑥, 𝜔), where 𝑥 represents a configuration we wish to optimize, partially controllable objective function,
and 𝜔 represents relevant parameters of the environment not under 𝑔 (𝑥, 𝜔)

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 277

 extensions and related settings

environmental parameter, 𝜔 our control. In this context, 𝜔 is called a environmental parameter or

special case: perturbed parameters environmental variable. A notable special case of this setup is where the
environmental parameters represent a perturbation to the optimization
parameters at the time of performance:

𝑔(𝑥, 𝜔) = ℎ(𝑥 + 𝜔), (11.22)

where ℎ is an objective function to be optimized as usual. Such a model

would allow us to consider scenarios where there may be instability or
imprecision in the specification of control parameters.
The primary challenge when facing such an objective is in identify-
ing a clear optimization goal in the face of parameters that cannot be
controlled; once this has been properly formalized, the development of
optimization policies is then usually relatively straightforward. Several
possible approaches have been studied to this end, which vary some-
what in their motivation and details. One detail with strong influence on
algorithm design is whether the environment and/or any perturbations
to inputs can be controlled during optimization; if so, we can accelerate
optimization through careful manipulation of the environment.

Stochastic optimization
81 This is a heavily overloaded phrase, as it is In stochastic optimization,81 we seek to optimize the expected performance
used as an umbrella term for optimization in- of the system of interest (also known as the integrated response82) under
volving any random elements. It is also used,
for example, to describe optimization from an assumed distribution for the environmental parameters, 𝑝 (𝜔):
noisy observations, to include methods such ∫
as stochastic gradient descent. As noisy ob- 𝑓 = 𝔼𝜔 [𝑔] = 𝑔(𝑥, 𝜔) 𝑝 (𝜔) d𝜔. (11.23)
servations are commonplace in Bayesian opti-
mization, we reserve the phrase for stochastic
environmental parameters only. Optimizing this objective presents a challenge: in general, the expec-
82 b. j. williams et al. (2000). Sequential Design tation with respect to 𝜔 cannot be evaluated directly but only estimated
of Computer Experiments to Minimize Inte- via repeated trials in different environments, and estimating the expected
grated Reseponse Functions. Statistica Sinica performance (11.23) with some degree of precision may require numerous
10(4):1133–1152.
evaluations of 𝑔(𝑥, 𝜔). However, when 𝑔 itself is expensive to evaluate
– for example, if every trial requires manual manipulation of a robot
and its environment before we can measure performance – this may not
be the most efficient approach, as we may waste significant resources
shoring up our belief about suboptimal values.
Instead, when we can control the environment during optimiza-
83 If 𝑔 has distribution GP (𝜇, 𝐾), then 𝑓 has dis- tion at will, we can gain some traction by designing a sequence of free
tribution GP (𝑚, 𝐶) with: parameter–environment pairs, potentially changing both configuration
∫
𝑚 = 𝜇 (𝑥, 𝜔) 𝑝 (𝜔) d𝜔; and environment in each iteration. The most direct way to design such
∬ an algorithm in our framework would be to model the environmental-


𝐶= 𝐾 [𝑥, 𝜔 ], [𝑥 ,0 𝜔 0 ] 𝑝 (𝜔, 𝜔 0 ) d𝜔 d𝜔 0. conditional objective function 𝑔 directly and define a utility function
and policy with respect to this function, in light of the environmental-
marginal objective 𝑓 (11.23) and its induced Gaussian process distribution.
Bayesian quadrature: § 2.6, p. 33 A Gaussian process model on 𝑔 is particularly practical in this regard,
as we may then use Bayesian quadrature to seamlessly estimate and
quantify our uncertainty in the objective 𝑓 (11.23) from observations.83

278
 11.9. stochastic and robust optimization

This approach was explored in depth by toscano-palmerin and frazier,

who also provided an excellent review of the related literature.84 84 s. toscano-palmerin and p. i. frazier (2018).
Several algorithms have also been proposed for the special case Bayesian Optimization with Expensive Inte-
grands. arXiv: 1803.08661 [stat.ML].
of perturbed control parameters (11.22), where we seek to optimize an
objective function ℎ under a random perturbation of its input, whose
distribution may depend on location:
∫
𝑓 = 𝔼𝜔 [ℎ] = ℎ(𝑥 + 𝜔) 𝑝 (𝜔 | 𝑥) d𝜔, (11.24)

In effect, the goal is to identify relatively “stable” (or “wide”) optima of 𝑥 𝑥0

the objective function; see the margin for an illustration. Once again, a When stability to perturbation is critical, the
Gaussian process belief on ℎ induces a Gaussian process belief on the relatively stable, but suboptimal, peak at 𝑥 0
averaged objective (11.24),85 which can aid in the construction of policies. might be preferred to the relatively narrow
When the perturbations follow a multivariate Gaussian distribution, global optimum at 𝑥.
and when inputs are assumed to be perturbed during optimization (and
85 If ℎ has distribution GP (𝜇, 𝐾), then 𝑓 has dis-
not only at performance time), one simple approach is to estimate the ex- tribution GP (𝑚, 𝐶) with:
pectation an arbitrary acquisition function 𝛼 (𝑥; D) (defined with respect ∫
to the unperturbed objective ℎ) with respect to the random perturbations, 𝑚 = 𝜇 (𝑥 + 𝜔) 𝑝 (𝜔 | 𝑥) d𝜔;
∫ ∬
𝐶= 𝐾 (𝑥 + 𝜔, 𝑥 0 + 𝜔 0 )
𝔼𝜔 [𝛼] = 𝛼 (𝑥 + 𝜔; D) 𝑝 (𝜔) d𝜔,
𝑝 (𝜔, 𝜔 0 | 𝑥, 𝑥 0 ) d𝜔 d𝜔 0.
via an unscented transform.86 We may then maximize the expected ac-
quisition function to realize an effective policy. 86 j. garcía-barcos and r. martinez-cantin
(2021). Robust policy search for robot naviga-
Under the assumption of unperturbed parameters during optimiza- tion. ieee Robotics and Automation Letters 6(2):
tion, fröhlich et al. described an adaptation of max-value entropy 2389–2396.
search for maximizing the expected objective (11.24).87 oliveira et al.
The authors focused on expected improve-
studied what is effectively the reversal of this problem: optimizing the ment and also defined an “unscented incum-
unperturbed objective ℎ (11.24) under the assumption of perturbed inputs bent” to approximate the maximum of (11.24).
during optimization. The authors proposed a variation of the Gaussian 87 l. p. fröhlich et al. (2020). Noisy-Input En-
Process upper confidence bound algorithm for this setting with theoreti- tropy Search for Efficient Robust Bayesian Op-
cal guarantees on convergence.88 timization. aistats 2020.
88 r. oliveira et al. (2019). Bayesian optimisation
under uncertain inputs. aistats 2019.
Robust optimization
Maximizing expected performance (11.23, 11.24) may not always be appro- risk tolerance: § 6.1, p. 111
priate in all settings, as it is risk neutral by design. In the interest of risk
aversion, in some scenarios we may instead seek parameters that ensure
reasonable performance even in unfavorable environments. This is the
goal of robust optimization. We can identify two families of approaches
to robust optimization, differing according to whether we seek guaran-
tees that are probabilistic (good performance in most environments) or
adversarial (good performance in even the worst possible environment).
𝜔
Probabilistic approaches focus on optimizing risk measures of per-
formance such as the value-at-risk or conditional value-at-risk. In the var

former case, we seek to optimize a (lower) quantile of the distribution of The value-at-risk (at the 𝜋 = 5% level) for an
𝑔(𝑥, 𝜔), rather than its mean (11.23); see the figure in the margin: example distribution of 𝑔 (𝑥, 𝜔). The var is

a pessimistic lower bound on performance
var(𝑥; 𝜋) = inf 𝛾 | Pr 𝑔(𝑥, 𝜔) ≤ 𝛾 | 𝑥 ≥ 𝜋 . holding with probability 1 − 𝜋 .

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 279

 extensions and related settings

We can interpret the value-at-risk as a pessimistic lower bound on

performance with tunable failure probability 𝜋. Although a natural mea-
sure of robustness, the value-at-risk ignores the shape of the upper tail
of the performance distribution entirely – any two distributions with
equal 𝜋-quantiles are judged equal. The conditional value-at-risk mea-
𝜔 sure accounts for differences in upper tail distributions by computing
var ∗ ∗ the expected performance in the favorable regime:
These distributions have the same value-at-
 
cvar(𝑥; 𝜋) = 𝔼𝜔 𝑔(𝑥, 𝜔) | 𝑔(𝑥; 𝜔) ≥ var(𝑥; 𝜋) .
risk at 𝜋 = 5%; however, the blue distribution
has greater conditional value-at-risk (marked
∗), as its upper tail has a higher expectation Two performance distributions that cannot be distinguished by var but
than the red distribution’s (marked ∗). can be distinguished by their cvar are illustrated in the marginal figure.
Optimizing both value at risk and conditional value at risk has re-
89 q. p. nguyen et al. (2021b). Value-at-Risk Opti- ceived some attention in the literature, and optimization policies for this
mization with Gaussian Processes. icml 2021. setting – under the assumption of a controllable environment during op-
90 q. p. nguyen et al. (2021a). Optimizing Condi- timization – have been proposed based on upper confidence bounds,89, 90
tional Value-At-Risk of Black-Box Functions.
Thompson sampling,90 and the knowledge gradient.91
neurips 2021.
An alternative approach to robust optimization is to maximize some
91 s. cakmak et al. (2020). Bayesian Optimization
of Risk Measures. neurips 2020. notion of worst-case performance. For example, we might consider an
adversarial objective function of the form:

𝑓 (𝑥) = min 𝑔(𝑥, 𝜔), (11.25)

𝜔 ∈Δ(𝑥)

where Δ(𝑥) is a compact, possibly location-dependent subset of environ-

ments to consider. Note that value-at-risk is a actually special case of
92 bogunovic et al.’s main focus was actually the this construction, where Δ(𝑥) is taken to be the upper tail of our belief
related problem of adversarial perturbations: over 𝑔(𝑥, 𝜔); however, we could also define such an adversarial objective
𝑓 (𝑥) = min ℎ (𝑥 + 𝜔); without any reference to probability at all. bogunovic et al. for example
considered robust optimization in this adversarial setting92 and described
𝜔 ∈Δ(𝑥 )

however, they also extend their algorithm to a simple algorithm based on upper/lower confidence bounds with strong
robust objectives of the form given in (11.25).
convergence guarantees.
kirschner et al. considered a related adversarial setting: distribution-
ally robust optimization, where we seek to optimize effectively without
perfect knowledge of the environment distribution. They considered an
objective of the following form:
∫
𝑓 (𝑥) = inf 𝑔(𝑥, 𝜔) 𝑝 (𝜔) d𝜔.
𝑝 ∈P

where P is a space of possible probability measures over the environment

to consider – thus we seek to maximize expected performance under the
reproducing kernel Hilbert spaces: § 10.2, worst-case environment distribution. The authors proposed an algorithm
p. 219 based on upper confidence bounds and proved convergence guarantees
under technical assumptions on 𝑔 (bounded rkhs norm) and P (bounded
93 r. martinez-cantin et al. (2018). Practical maximum mean discrepancy from a reference distribution).
Bayesian optimization in the presence of out-
liers. aistats 2018.
Finally, another type of robustness that has been received some
94 i. bogunovic et al. (2020). Corruption-
attention is robustness to extreme measurement errors (beyond additive
Tolerant Gaussian Process Bandit Optimiza- Gaussian noise), including heavy-tailed noise such as Student-𝑡 errors
tion. aistats 2020. (2.31)93 and adversarial corruption.94

280
 11.10. incremental optimization of seqential procedures

11.10 incremental optimization of seqential procedures

In some applications, the objective function is determined by a sequence
of dependent steps eventually producing its final value. If we have access
to this sequential process and can model its progression, we may be
able to accelerate optimization via shrewd “early stopping”: terminating
evaluations still in progress when their final value can be forecasted with
sufficient confidence.
Hyperparameter tuning presents one compelling example. Consider
for example the optimization of neural network hyperparameters 𝜽 .95 95 We will use 𝜽 for the variable to be optimized
The objective function in this setting is usually defined to be the value of here rather than 𝑥 to be consistent with our
previous discussion of Gaussian process hy-
some loss function ℓ (for example, validation error) after the network has perparameters in chapter 3.
been trained with the chosen hyperparameters. However, this training
is an iterative procedure: if the network is parameterized by a vector of
weights w, then the objective function might be defined by

𝑓 (𝜽 ) = lim ℓ (w𝑡 ; 𝜽 ), (11.26)

𝑡 →∞

the loss of the network after the weights have converged.96 If we don’t 96 One could consider early stopping in the train-
treat this objective function as a black box but take a peek inside, we ing procedure as well, but this simple example
may interpret it as the limiting value of the learning curve defined by the is useful for exposition.

learning procedure’s loss at each stage of training.

ℓ
The iterative nature of this objective function offers the opportunity
for innovation in policy design. Learning curves typically exhibit fairly
regular behavior, with the loss in each step of training generally falling
over time until settling on its final value. This suggests we may be able
to faithfully extrapolate the final value of a learning curve from the
early stages of training; a toy example is presented in the margin. When
possible, we may then be able to speed up optimization by not always
𝑡
training to convergence with every setting of the hyperparameters we
explore. With this motivation in mind, several sophisticated methods for Learning curve extrapolation from initial
extrapolating learning curves have been developed, including carefully training results. At this point, we may already
crafted parametric models97, 98 and flexible Bayesian neural networks.99 wish to abandon the red option.
Exploiting the ability to extrapolate sequential objectives requires
that we expand our action space (step 1) to allow fine-grained control over 97 k. swersky et al. (2014). Freeze–Thaw
evaluation. One compelling scheme was proposed by swersky et al.,97 Bayesian Optimization. arXiv: 1406 . 3896
who suggested maintaining a set of partial evaluations of the objective [stat.ML].
throughout optimization. In the context of our hyperparameter tuning 98 t. domhan et al. (2015). Speeding up Auto-
matic Hyperparameter Optimization of Deep
example (11.26), we would maintain a set of thus-far investigated hyper-
Neural Networks by Extrapolation of Learning
parameters {𝜽𝑖 }, each accompanied by a sequence (of variable length) of Curves. ijcai 2015.
thus-far evaluated weights {w𝑡,𝑖 } and the associated losses {ℓ𝑡,𝑖 }. Now, 99 a. klein et al. (2017). Learning Curve Pre-
each action we design can either investigate a novel hyperparameter 𝜽 diction with Bayesian Neural Networks. iclr
or extend an existing partial evaluation by one step. The authors dubbed 2017.
this scheme freeze–thaw Bayesian optimization, as after each action we
“freeze” the current evaluation, saving enough state such that we can freeze–thaw Bayesian optimization
“thaw” it later for further investigation if desired.
Regardless of modeling details, this scheme offers the potential for
considerable savings when optimizing sequential objective functions.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 281

 extensions and related settings

100 l. li et al. (2018b). Hyperband: A Novel Bandit- This idea of abandoning stragglers based on early progress is also the
Based Approach to Hyperparameter Optimiza- basis of the bandit-based hyperband algorithm.100
tion. Journal of Machine Learning Research
18(185):1–52.
11.11 non-gaussian observation models and active search
Throughout this book, we have focused almost exclusively on the additive
Gaussian noise observation model. There are good reasons for this: it is
a reasonably faithful model of many systems and offers exact inference
with Gaussian process models of the objective function. However, the
assumption of Gaussian noise is not always warranted and may be
fundamentally incompatible with some scenarios.
decision theory for optimization: chapter 5, Fortunately, the decision-theoretic core of most Bayesian optimiza-
p. 87 tion approaches does not make any assumptions regarding our model
of the objective function or our observations of it, and with some care
utility functions for optimization: chapter 6, the utility functions we developed for optimization can be adapted for
p. 109 virtually any scenario. Further, there are readily available pathways for
gp inference with non-Gaussian observation incorporating non-Gaussian observation models into Gaussian process
models: § 2.8, p. 35 objective function models, so we do not need to abandon that rich model
class in order to use alternatives.

Sequential optimization with non-Gaussian observation models

A decision-theoretic approach to optimization entails first selecting an
objective function model 𝑝 (𝑓 ) and observation model 𝑝 (𝑦 | 𝑥, 𝜙), which
together are sufficient to derive the predictive distribution 𝑝 (𝑦 | 𝑥, D)
relevant to every sequential decision made during optimization. After
selecting a utility function 𝑢 (D), we may then follow the iterative pro-
cedure developed in chapter 5 to derive a policy.
101 a. shah et al. (2014). Student-𝑡 Processes as This abstract approach has been realized in several specific settings.
Alternatives to Gaussian Processes. aistats For example, both Student-𝑡 processes101 and the Student-𝑡 observation
2014.
model102 have been explored to develop Bayesian optimization routines
102 r. martinez-cantin et al. (2018). Practical that are robust to the presence of outliers.103
Bayesian optimization in the presence of out-
liers. aistats 2018. tesch et al. explored the use of expected improvement for optimiza-
103 We explored this possibility at length in § 2.8. tion from binary success/failure indicators, motivated by optimizing the
probability of success of a robotic platform operating in an uncertain
104 m. tesch et al. (2013). Expensive Function Op- environment.104 Here the utility function was taken to be this success
timization with Stochastic Binary Outcomes. probability maximized over the observed locations:
icml 2013.

𝑢 (D) = max 𝔼[𝑦 | 𝑥, D] = max Pr(𝑦 = 1 | 𝑥, D),

x x

105 Note that the simple reward utility function where for this expression we have assumed binary outcomes 𝑦 ∈ {0, 1}
(6.3) we have been working with can be writ- with 𝑦 = 1 interpreted as indicating success.105 The authors then derived
ten in this exact form if we assume any addi-
tive observation noise has zero mean.
a policy for this setting via one-step lookahead.
Another setting involving binary (or categorical) feedback is in opti-
106 n. houlsby et al. (2012). Collaborative
Gaussian Processes for Preference Learning. mizing human preferences, such as in a/b testing or user modeling. Here
neurips 2012. we might seek to optimize user preferences by repeatedly presenting a
panel of options and asking for the most preferred item. houlsby et al. de-
scribed a convenient reduction from preference learning to classification

282
 11.11. non-gaussian observation models and active search

for Gaussian processes that allows the immediate use of standard poli-
cies such as expected improvement,106, 104 although more sophisticated
policies have also been proposed specifically for this setting.107 107 See appendix d, p. 329 for related references.

Active search
garnett et al. introduced active search as a simple model of scientific
discovery in a discrete domain X = {𝑥𝑖 }.108 In active search, we assume 108 r. garnett et al. (2012). Bayesian Optimal Ac-
that among these points is hidden a rare, valuable subset exhibiting tive Search and Surveying. icml 2012.
desirable properties for the task at hand. Given access to an oracle that
can – at significant cost – determine whether an identified point belongs
to the sought after class, the problem of active search is to design a
sequence of experiments seeking to maximize the number of discoveries
in a given budget. A motivating application is drug discovery, where
the domain would represent a list of candidate molecules to search for
those rare examples exhibiting significant binding activity with a chosen
biological target. As the space of candidates is expansive and the cost of virtual screening: appendix d, p. 314
even virtual screening is nontrivial, intelligent experimental design has
the potential to greatly improve the rate of discovery.
To derive an active search policy in our framework, we must first modeling observations
model the observation process and determine a suitable utility function.
The former requires consideration of the nuances of a given situation,
but we may provide a barebones construction that is already sufficient
to be of practical and theoretical interest. Given a discrete domain X,
we assume there is some identifiable subset V ⊂ X of valuable points
we wish to recover. We associate with each point 𝑥 ∈ X a binary label 109 Other situations may call for other approaches;
𝑦 = [𝑥 ∈ V] indicating whether 𝑥 is valuable (𝑦 = 1) or not (𝑦 = 0). A for example, if value is determined by thresh-
olding a continuous measurement, we may
natural observation model is then to assume that selecting a point 𝑥 wish to model that continuous observation
for investigation reveals this binary label 𝑦 in response.109 Finally, we process explicitly.
may define a natural utility function for active search by assuming that,
all other things being held equal, we prefer a dataset containing more
valuable points to one with fewer: utility function
∑︁
𝑢 (D) = 𝑦. (11.27)
𝑥 ∈D

This is simply the cumulative reward utility (6.7), which here can be 110 The assumption of a discrete domain is to
interpreted as counting the number of valuable points discovered.110 avoid repeatedly observing effectively the
same point to trivially “max out” this score.
To proceed with the Bayesian decision-theoretic approach, we must
build a model for the uncertain elements inherent to each decision. Here
the primary object of interest is the predictive posterior distribution posterior predictive probability,
Pr(𝑦 = 1 | 𝑥, D), the posterior probability that a given point 𝑥 is valuable. Pr(𝑦 = 1 | 𝑥, D)
We may build this model in any number of ways, for example by com-
bining a Gaussian process prior on a latent function with an appropriate 111 c. e. rasmussen and c. k. i. williams (2006).
choice of observation model.111 Gaussian Processes for Machine Learning. mit
Press. [chapter 3]
Equipped with a predictive model, deriving the optimal policy is a
simple exercise. To begin, the one-step marginal gain in utility (5.8) is

𝛼 1 (𝑥; D) = Pr(𝑦 = 1 | 𝑥, D);

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 283

 extensions and related settings

that is, the optimal one-step decision is to greedily maximize the proba-
bility of success. Although this is a simple (and somewhat obvious) policy
that can perform well in practice, theoretical and empirical study on
112 r. garnett et al. (2012). Bayesian Optimal Ac- active search has established that massive gains can be had by adopting
tive Search and Surveying. icml 2012. less myopic policies.
113 s. jiang et al. (2017). Efficient Nonmyopic Ac- On the theoretical side, garnett et al. demonstrated by construc-
tive Search. icml 2017. tion that the expected performance of any lookahead approximation
114 r. garnett et al. (2015). Introducing the ‘ac- can be exceeded by any arbitrary amount by extending the lookahead
tive search’ method for iterative virtual screen- horizon even a single step.112 This result was strengthened by jiang
ing. Journal of Computer-Aided Molecular De-
sign 29(4):305–314. et al., who showed – again by construction – that no policy that can be
computed in time polynomial in |X | can approximate the performance of
the optimal policy within any constant factor.113 Thus the optimal active
cost of computing optimal policy: § 5.3, p. 99 search policy is not only hard to compute, but also hard to approximate.
These theoretical results, which rely on somewhat unnatural adversarial
constructions, have been supported by empirical investigations on real-
world data as well.113, 114 For example, garnett et al. demonstrated that
simply using two-step instead of one-step lookahead can significantly
accelerate virtual screening for drug discovery across a broad range of
biological targets.114
This is perhaps a surprising state of affairs given the success of
one-step lookahead – and the relative lack of less myopic alternatives –
for traditional optimization. We can resolve this discrepancy by noting
moving beyond one-step lookahead: § 7.10, that utility functions used in that setting, such as the simple reward
p. 150 (6.3), inherently exhibit decreasing marginal gains as they are effectively
bounded by the global maximum. Further, such a utility tends to remain
relatively constant throughout optimization, punctuated by brief but
significant increases when a new local maximum is discovered. On the
other hand, for the cumulative reward (11.27), every observation has
the potential to increase the utility by exactly one unit. As a result,
every observation is on equal footing in terms of potential impact, and
there is increased pressure to consider the entire search trajectory when
designing each observation.
One thread of research on active search has focused on developing
efficient, yet nonmyopic policies grounded in approximate dynamic pro-
gramming, such as lookahead beyond one step.112 jiang et al. proposed
batch rollout: § 5.3, p. 103 one significantly less myopic alternative policy based on batch rollout.113
The key observation is that we may construct the one-step optimal batch
observation of size 𝑘 by computing the posterior predictive probability
Pr(𝑦 = 1 | 𝑥, D) for the unlabeled points, sorting, and taking the top 𝑘;
this is a consequence of linearity of expectation and utility (11.27). With
this, we may realize an efficient batch rollout policy for horizon 𝜏 by
maximizing the acquisition function
 ∑︁ 
0
Pr(𝑦 = 1 | 𝑥, D) + 𝔼𝑦 Pr(𝑦 = 1 | 𝑥 , D1 ) | 𝑥, D .
0 0
(11.28)
𝜏−1
Í0
Here the sum-with-prime notation 𝜏−1 indicates the sum of the top-
(𝜏 −1) values over the unlabeled data – the expected utility of the optimal

284
 11.11. non-gaussian observation models and active search

Figure 11.17: A demonstration of active

search. Left: a 2𝑑 domain col-
ored according to the prob-
ability of a positive result.
Right: 500 points chosen by
a nonmyopic active search
policy reflecting considera-
tion of exploration versus ex-
ploitation. The observations
colored red were positive
and those colored gray were
negative.

batch observation consuming the remaining budget.115 In experiments,

115 The cost of computing this acquisition func-
this policy showed interesting emergent behavior: it underperforms looka- tion is O (𝑛 2 log 𝑛), where 𝑛 = |X |; this is
head policies in the early stages of search due to significant investment roughly the same order as the two-step ex-
in early exploration. However, this exploration pays off dramatically pected marginal gain, O (𝑛 2 ).
by termination by revealing promising numerous regions that may be
exploited later.
Figure 11.17 illustrates active search in a 2𝑑 domain, which for the
purposes of this demonstration was discretized into a 100 × 100 grid.
The example is constructed so that a small number of discrete regions
yield valuable items, which must be efficiently uncovered for a successful
search. The right-hand panel shows a sequence of 500 observations de-
signed by iteratively maximizing (11.28); their distribution clearly reflects
both exploration of the domain and exploitation of the most fruitful
regions. The rate of discovery for the active search policy was approxi-
mately 3.8 times greater than would be expected from random search.
vanchinathan et al. considered an expanded setting they dubbed
adaptive valuable item discovery (avid), for which active search is a
special case. Here items may have nonnegative, continuous values and
the cumulative reward utility (11.27) was augmented with a term en-
couraging diversity among the selected items.116 The authors proposed 116 h. p. vanchinathan et al. (2015). Discovering
an algorithm called gp-select for this setting based on an acquisition Valuable Items from Massive Data. kdd 2015.
function featuring two terms: a standard upper confidence bound score 117 a. kulesza and b. taskar (2012). Determi-
(8.25) and a term encouraging diversity related to determinantal point nantal Point Processes for Machine Learning.
Foundations and Trends in Machine Learning
processes.117 The authors were further able to establish theoretical regret 5(2–3):123–286.
bounds for this algorithm under standard assumptions on the complexity
of the value function.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 285

 12
A BRIEF HISTORY OF BAYESIAN OPTIMIZATION

In this chapter we provide a historical survey of the ideas underpinning

Bayesian optimization, including important mathematical precedents.
We also document the progression of major ideas in Bayesian optimiza-
tion, from its first appearance in 1962 to the present day. A major goal will
be to identify the point of introduction of prominent Bayesian optimiza-
tion policies, as well as notable instances of subsequent reintroduction
when ideas were forgotten and later rediscovered and refined.

12.1 historical precursors and optimal design

The Bayesian approach to optimization is founded on a simple premise:
experiments should be designed with purpose, both guided by our knowl-
edge and aware of our ignorance. The optimization policies built on this
principle can be understood as statistical manifestations of rational in-
quiry, where we design a sequence of experiments to systematically
reveal the maximal value attained by the system of interest.
Statistical approaches to experimental design have a long history,
with the earliest examples appearing over 200 years ago.1, 2 A landmark 1 j. d. gergonne (1815). Application de la méth-
early contribution was smith’s 1918 dissertation,3 which considered ode des moindres quarrés à l’interpolation des
suites. Annales de Mathématiques pures et ap-
experimental design for polynomial regression models to minimize a pliquées 6:242–252.
measure of predictive uncertainty. Shortly thereafter, fisher published
2 c. s. peirce (1876). Note on the Theory of the
a hugely influential guide to statistical experimental design based on Economy of Research. In: Report of the Superin-
his experience analyzing crop experiments at Rothamsted Experimental tendent of the United States Coast Survey Show-
Station,4 which was instrumental in shaping modern statistical practice. ing the Progress of the Work for the Fiscal Year
These works served to establish the field of optimal design, an expansive Ending with June, 1876.

subject which has now enjoyed a century of study. Numerous excellent 3 k. smith (1918). On the Standard Deviations
references are available.5, 6 of Adjusted and Interpolated Values of an Ob-
served Polynomial Function and its Constants
Early work in optimal design did not consider the possibility of and the Guidance they Give Towards a Proper
adaptively designing a sequence of experiments, an essential feature of Choice of the Distribution of Observations.
Bayesian optimization. Instead, the focus was optimizing fixed designs Biometrika 12(1–2):1–85.
to minimize some measure of uncertainty when performing inference 4 r. a. fisher (1935). The Design of Experiments.
with the resulting data. This paradigm is practical when experiments Oliver and Boyd.
are extremely time consuming but can easily run in parallel, such as 5 g. e. p. box et al. (2005). Statistics for Exper-
the agricultural experiments studied extensively by fisher. The most imenters: Design, Innovation, and Discovery.
John Wiley & Sons.
common goals considered in classical optimal design are accurate estima-
tion of model parameters and confident prediction at unseen locations; 6 d. c. montgomery (2019). Design and Analysis
of Experiments. John Wiley & Sons.
these goals are usually formulated in terms of optimizing some statisti-
cal criterion as a function of the design.7 However, in 1941, hotelling 7 These criteria often have alphabetic names:
notably studied experimental designs for estimating the location of the a-optimality, d-optimality, v-optimality, etc.
maximum of an unknown function in this nonadaptive setting.8 This 8 h. hotelling (1941). Experimental Determina-
was perhaps the first rigorous treatment of batch optimization. tion of the Maximum of a Function. The Annals
of Mathematical Statistics 12(1):20–45.

12.2 seqential analysis and bayesian experimental design

Concentrated study of sequential experiments began during World War
ii with wald, who pioneered the field of sequential analysis. The seminal

This material will be published by Cambridge University Press as Bayesian Optimization. 287
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 a brief history of bayesian optimization

result was a hypothesis testing procedure for sequentially gathered

data that can terminate dynamically once the result is known with
sufficient confidence. Such sequential tests can be significantly more data
efficient than tests requiring an a priori fixed sample size. The potential
benefit of wald’s work to the war effort was immediately recognized,
9 a. wald (1945). Sequential Tests of Statisti- and his research was classified and only published after the war.9, 10
cal Hypotheses. The Annals of Mathematical The introduction of sequential analysis would kick off multiple parallel
Statistics 16(2):117–186.
lines of investigation, leading to both Bayesian sequential experimental
10 a. wald (1947). Sequential Analysis. John Wiley design and multi-armed bandits, and eventually all modern approaches
& Sons.
to Bayesian optimization.
The success of sequential analysis lead numerous researchers to in-
vestigate sequential experimental design in the following years. Sequential
experimental design for optimization was proposed by friedman and
11 m. friedman and l. j. savage (1947). Plan- savage as early as 1947.11 They argued that nonadaptive optimization pro-
ning Experiments Seeking Maxima. In: Se- cedures, as considered earlier by hotelling,12 can be wasteful as many
lected Techniques of Statistical Analysis for Sci-
entific and Industrial Research, and Production
experiments may be squandered needlessly exploring suboptimal re-
and Management Engineering. gions. Instead, friedman and savage suggested sequential optimization
12 h. hotelling (1941). Experimental Determina- could be significantly more efficient, as poor regions of the domain can
tion of the Maximum of a Function. The Annals be quickly discarded while more time is spent exploiting more promising
of Mathematical Statistics 12(1):20–45. areas. Their proposed algorithm was a simple procedure entailing suc-
cessive axis-aligned line searches, optimizing each input variable in turn
while holding the others fixed – what would now be known as cyclic
coordinate descent.
13 g. e. p. box and k. b. wilson (1951). On the Ex- box greatly expounded on these ideas, working for years alongside
perimental Attainment of Optimum Condi- a chemist (wilson) experimentally optimizing chemical processes as a
tions. Journal of the Royal Statistical Society function of environmental and process parameters, for example, opti-
Series B (Methodological) 13(1):1–45.
mizing product yield as a function of reactant concentrations.13, 14 box
14 g. e. p. box (1954). The Exploration and Ex- advocated the method of steepest ascent during early stages of opti-
ploitation of Response Surfaces: Some General
Considerations and Examples. Biometrics 10(1):
mization rather than the “one factor at a time” heuristic described by
16–60. friedman and savage, pointing out that the latter method is prone to
15 g. e. p. box and p. v. youle (1954). The Explo- becoming stuck in ridge-shaped features of the objective. box and youle
ration and Exploitation of Response Surfaces: also provided insightful commentary on how the process of optimization,
An Example of the Link Between the Fitted and in particular geometric features of the objective function surface,
Surface and the Basic Mechanism of the Sys-
tem. Biometrics 11(3):287–323.
may lead the experimenter to a greater fundamental understanding of
underlying physical processes.15
16 h. chernoff (1959). Sequential Design of Ex-
periments. The Annals of Mathematical Statis-
In the following years, researchers developed general methods for
tics 30(3):755–770. sequential experimental design targeting a broad range of experimen-
17 h. chernoff (1972). Sequential Analysis and tal goals. An early pioneer was chernoff, a student of wald’s, who
Optimal Design. Society for Industrial and Ap- extended sequential analysis to the adaptive setting and provided asymp-
plied Mathematics. totically optimal procedures for sequential hypothesis testing.16 He also
18 j. bather (1996). A Conversation with Her- wrote an survey of this early work,17 and would eventually remark in an
man Chernoff. Statistical Science 11(4):335–350. interview:18
Although I regard myself as non-Bayesian, I feel in sequential prob-
lems it is rather dangerous to play around with non-Bayesian pro-
cedures.
Another important contribution around this time was the reintroduc-
tion of multi-armed bandits by robbins, which would quickly explode

288
 12.3. the rise of bayesian optimization

into a massive body of literature. We will return to this line of work

momentarily.
The Bayesian approach to sequential experimental design was for-
malized shortly after chernoff’s initial work. Authors such as raiffa
and schlaifer19 and lindley20 promoted a general approach based on 19 h. raiffa and r. schlaifer (1961). Applied Sta-
Bayesian decision theory, wherein the experimenter selects a utility tistical Decision Theory. Division of Research,
Graduate School of Business Administration,
function reflecting their experimental goals and a model for reasoning Harvard University.
about experimental outcomes given data, then designs each experiment
20 d. v. lindley (1972). Bayesian Statistics, A Re-
to maximize the expected utility of the collected data. This is precisely view. Society for Industrial and Applied Math-
the procedure we outlined in chapter 5. As we noted in our presentation, ematics.
this framework yields theoretically optimal policies, but comes with an
unwieldy computational burden.

12.3 the rise of bayesian optimization

By the early 1960s, the stage was set for Bayesian optimization, which
could now be realized by appropriately adapting the now mature field of
Bayesian experimental design.
kushner was the first to seize the opportunity with a pair of papers
on optimizing a one-dimensional objective observed with noise.21, 22 All 21 h. j. kushner (1962). A Versatile Stochastic
of the major ideas in modern Bayesian optimization were already in Model of a Function of Unknown and Time
Varying Form. Journal of Mathematical Analy-
place in this initial work, including a Gaussian process model of the sis and Applications 5(1):150–167.
objective function and appealing to Bayesian decision theory to derive
22 h. j. kushner (1964). A New Method of Locat-
optimization policies. After dismissing the optimal policy (with respect ing the Maximum Point of an Arbitrary Multi-
to the global reward utility (6.5)) as “notoriously difficult”21 and “virtu- peak Curve in the Presence of Noise. Journal
ally impossible to compute,”22 kushner suggests two alternatives “on of Basic Engineering 86(1):97–106.
the basis of heuristic or intuitive considerations”: maximizing an upper
confidence bound (§ 7.8) and maximizing probability of improvement
(§ 7.5), which share credit as the first Bayesian optimization policies to ap-
pear in the literature.21 The probability of improvement approach seems
to have won his favor, and kushner later provided extensive practical
advice for realizing this method, including careful discussion of how
the improvement threshold could be managed interactively throughout
optimization by a human expert in the loop.22
A significant body of literature on Bayesian optimization emerged
in the Soviet Union following kushner’s seminal work. Many of these
authors notably explored the promise of one-step lookahead for effective
policy design, proposing and studying both expected improvement and
the knowledge gradient for the first time. šaltenis23 was the first to 23 Also transliterated shaltyanis.
introduce expected improvement (§ 7.3) in 1971.24 This work contains 24 v. r. šaltenis (1971). One Method of Multiex-
an explicit formula for expected improvement for arbitrary Gaussian tremum Optimization. Avtomatika i Vychisli-
process models and the results of an impressive empirical investigation tel’naya Tekhnika (Automatic Control and Com-
puter Sciences) 5(3):33–38.
on a Soviet mainframe computer with objective functions in dimensions
up to 32. šaltenis concludes with the following observation:

The relatively large amounts of machine time spent in planning

search and the complexity of the algorithm give us grounds to
assume that the most effective sphere of application would be mul-

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 289

 a brief history of bayesian optimization

tiextremum target functions whose determination involves major

computational difficulties.

This remains the target domain of Bayesian optimization today.

25 Often transliterated močkus in early work. Another prominent early contributor was mockus,25 who wrote a
26 j. mockus (1972). Bayesian Methods of Search series of papers on Bayesian optimization in the 1970s26, 27, 28 and has
for an Extremum. Avtomatika i Vychisli- written two books on the subject.29, 30 Like kushner, mockus begins his
tel’naya Tekhnika (Automatic Control and Com- presentation in these papers by outlining the optimal policy for maxi-
puter Sciences) 6(3):53–62.
mizing the global reward utility (6.5), but rejects it as computationally
27 j. mockus (1974). On Bayesian Methods for infeasible. As a practical alternative, mockus instead promotes what he
Seeking the Extremum. Optimization Tech-
niques ifip Technical Conference. calls the “one-stage approach,” that is, one-step lookahead.26 As he had
chosen the global reward utility, the resulting optimization policy is to
28 j. mockus et al. (1978). The Application of
Bayesian Methods for Seeking the Extreme- maximize the knowledge gradient (§ 7.4).
mum. In: Towards Global Optimization 2. This claim may give some readers pause, as mockus’s work is fre-
29 j. mockus (1989). Bayesian Approach to Global quently cited as the origin of expected improvement instead. However,
Optimization: Theory and Applications. Kluwer this is inaccurate for multiple reasons. Expected improvement had been
Academic Publishers. introduced by šaltenis in 1971, one year before mockus’s first contri-
30 j. mockus et al. (2010). Bayesian Heuristic Ap- bution on Bayesian optimization. Indeed, mockus was aware of and
proach to Discrete and Global Optimization: Al- cited šaltenis’s work. Further, the acquisition function mockus de-
gorithms, Visualization, Software, and Applica-
tions. Kluwer Academic Publishers. scribes is defined with respect to the global reward utility underlying the
31 See equation 8 in citation 26 above.
knowledge gradient,31 not the simple reward utility underlying expected
improvement.
That said, the situation is slightly more subtle. mockus also dis-
cusses two convenient choices of models on the unit interval for which
the knowledge gradient happens to equal expected improvement: the
Wiener process and the Ornstein–Uhlenbeck (ou) process. Both are rare
examples of Gauss–Markov processes, whose Markovian property ren-
ders sequential inference particularly convenient, and the early work
on Bayesian optimization was dominated by these models due to the
32 j. mockus (1972). Bayesian Methods of Search extreme computational limitations at the time. mockus also points out
for an Extremum. Avtomatika i Vychisli- these models have a “special propert[y]”;32 namely, their Markovian na-
tel’naya Tekhnika (Automatic Control and Com-
puter Sciences) 6(3):53–62. [equations 36–37]
ture ensures that the posterior mean is always maximized at an observed
location, and thus the simple and global reward utilities coincide!

12.4 later rediscovery and development

The expected improvement and the knowledge gradient acquisition
strategies were both reintroduced decades later when computational
power had increased to the point that Bayesian optimization could be a
practical approach for real problems.
33 m. schonlau (1997). Computer Experiments schonlau33 and jones et al.,34 working together, proposed maxi-
and Global Optimization. Ph.D. thesis. Univer-
sity of Waterloo. mizing expected improvement for efficient global optimization in the
context of the design and analysis of computer experiments (dace).35 Here
34 d. r. jones et al. (1998). Efficient Global Opti-
mization of Expensive Black-Box Functions. the objective function represents the output of a computational routine
Journal of Global Optimization 13(4):455–492. and is assumed to be observed without noise. In addition to promoting
35 j. sacks et al. (1989). Design and Analysis of expected improvement as a policy, jones et al. also provided extensive
Computer Experiments. Statistical Science 4(4): detail for practical implementation, including an insightful discussion on
409–435. model validation and a branch-and-bound strategy for maximizing the

290
 12.4. later rediscovery and development

expected improvement acquisition function for a certain class of Gaus-

sian process models. The knowledge gradient was picked up again and 36 p. frazier and w. powell (2007). The Knowl-
edge Gradient Policy for Offline Learning with
further developed by frazier and powell,36 who coined the name and Independent Normal Rewards. adprl 2007.
studied the policy in the discrete and independent (bandit-like) setting,
37 w. scott et al. (2011). The Correlated Knowl-
and scott et al.,37 who adopted the policy for continuous optimization. edge Gradient for Simulation Optimization of
Three years after reintroducting expected improvement, jones wrote Continuous Parameters Using Gaussian Pro-
an extensive survey of then-current Bayesian optimization policies.38 cess Regression. siam Journal on Optimization
Despite the now-pervasive nature of expected improvement, it is striking 21(3):996–1026.
that jones actually promotes maximizing the probability of improvement 38 d. r. jones (2001). A Taxonomy of Global Op-
as the most promising policy for Bayesian optimization throughout this timization Methods Based on Response Sur-
faces. Journal of Global Optimization 21(4):345–
survey. His concern with expected improvement was a potential lack of 383.
robustness if the objective function model is misspecified. His proposed
39 See § 7.5, p. 134 for a discussion of this pro-
alternative was maximizing the probability of improvement over a wide posal.
range of improvement targets and evaluating the objective function in 40 c. e. shannon (1948). A Mathematical Theory
parallel,39 which he regarded as more robust when practical. of Communication. The Bell System Technical
The concept of mutual information (§ 7.6) first appeared with shan- Journal 27(3):379–423.
non’s introduction of information theory,40 where it was called the This was a central concept in § 10.3, p. 222 as
channel capacity and served as a measure of the amount of informa- well.
tion that could be transferred effectively over a noisy communication 41 d. v. lindley (1956). On a Measure of the Infor-
channel. lindley later reinterpreted mutual information as the expected mation Provided by an Experiment. The An-
information gained by a proposed experiment and suggested maximizing nals of Mathematical Statistics 27(4):986–1005.
this quantity as an effective means of general Bayesian experimental 42 j. villemonteix et al. (2009). An informa-
design.41 tional approach to the global optimization
of expensive-to-evaluate functions. Journal of
The application of this information-theoretic framework to Bayes- Global Optimization 44(4):509–534.
ian optimization was first proposed by villemonteix et al.42 and later
43 p. hennig and c. j. schuler (2012). Entropy
independently by hennig and schuler,43 the latter of whom coined the Search for Information-Efficient Global Op-
now-prominent term entropy search. Both of these initial investigations timization. Journal of Machine Learning Re-
considered maximizing the mutual information between a measurement search 13(Jun):1809–1837.
and the location of the global optimum 𝑥 ∗ (§ 7.6), using the formula- 44 m. w. hoffman and z. ghahramani (2015).
tion in (7.14). hernández-lobato et al. later proposed a different set Output-Space Predictive Entropy Search for
of approximations based on the equivalent formulation in (7.15) under Flexible Global Optimization. Bayesian Opti-
mization Workshop, neurips 2015.
the name predictive entropy search, as the key quantity is the expected
reduction in entropy for the predictive distribution.60 Both hoffman 45 z. wang and s. jegelka (2017). Max-value En-
tropy Search for Efficient Bayesian Optimiza-
and ghahramani44 and wang and jegelka45 later pursued maximizing tion. icml 2017.
the mutual information with the value of the global optimum 𝑓 ∗ (§ 7.6).
46 j. m. calvin (1993). Consistency of a My-
These developments occurred contemporaneously and independently. opic Bayesian Algorithm for One-Dimensional
Prior to these algorithms designed to reduce the entropy of the value Global Optimization. Journal of Global Opti-
of the global maximum, there were occasional efforts to minimize some mization 3(2):223–232.
other measure of dispersion in this quantity. For example, calvin studied 47 In fact, the joint density 𝑝 (𝑥 ,∗ 𝑓 ∗ | D) is ana-
an algorithm for optimization on the unit interval X = [0, 1] wherein lytically tractable for the Wiener process with
drift, which forms the posterior on each subin-
the variance of 𝑝 (𝑓 ∗ | D) was greedily minimized via one-step looka- terval subdivided by the data:
head.46 Here the model was again the Wiener process, which has the
remarkable property that the distribution of 𝑝 (𝑓 ∗ | D) (and its variance) l. a. shepp (1979). The Joint Density of the Max-
imum and its Location for a Wiener Process
is analytically tractable.47 The Wiener and ou processes are among the with Drift. Journal of Applied Probability 16(2):
only nontrivial Gaussian processes with this property.48 423–427.
No history of Bayesian optimization would be complete without 48 r. j. adler and j. e. taylor (2007). Random
mentioning the role hyperparameter tuning has played in driving its Fields and Geometry. Springer–Verlag. [chap-
recent development. With the advent of deep learning, the early 2010s ter 4, footnotes 1–2]

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 291

 a brief history of bayesian optimization

saw the rise of extraordinarily complex learning algorithms trained on

extraordinarily large datasets. The great expense of training these models
created unprecedented demand for efficient hyperparameter tuning to
fuel the rapid development in the area. One could not ask for a more-
perfect fit for Bayesian optimization!
Interest in Bayesian optimization for hyperparameter tuning was
kicked off in earnest by snoek et al., who reported a dramatic improve-
ment in performance when tuning a convolutional neural network via
Bayesian optimization, even compared with carefully hand-tuned hyper-
49 j. snoek et al. (2012). Practical Bayesian Op- parameters.49 This served as a watershed moment for Bayesian optimiza-
timization of Machine Learning Algorithms. tion, leading to an explosion of interest and sustained effort from the
neurips 2012.
machine learning in the following years – although this history began
in 1815, over half of the works cited in this book are from after 2012!

12.5 multi-armed bandits to infinite-armed bandits

We have now covered the evolution of decision-theoretic Bayesian op-
timization policies from wald’s introduction of sequential analysis in
1945 to the state-of-the-art. Alongside these developments, a rich and
expansive body of literature was forming on the multi-armed problem. A
complete survey of this work would be out of this book’s scope; however,
50 d. a. berry and b. fristedt (1985). Bandit we can point the interested reader to comprehensive surveys.50, 51 Both
Problems: Sequential Allocation of Experiments. contain excellent bibliographic notes and combined serve as a indispens-
Chapman & Hall.
able guide to the literature on multi-armed bandits (§ 7.7). Our goal in
51 t. lattimore and c. szepesvári (2020). Bandit the following will be to cover developments that directly influenced the
Algorithms. Cambridge University Press.
evolution of Bayesian optimization.
thompson was the first to seriously study the possibility of sequen-
52 w. r. thompson (1933). On the Likelihood tial experiments in the context of medical treatment;52, 53 this work pre-
that One Unknown Probability Exceeds An- dated wald’s work by a decade. thompson considered a model scenario
other in View of the Evidence of Two Samples.
Biometrika 25(3–4):285–294.
where there are two possible treatments a doctor may prescribe for
a disease, but it is not clear which should be preferred. To determine
53 w. r. thompson (1935). On the Theory of Ap-
portionment. American Journal of Mathemat- the better treatment, we must undertake a clinical trial, assigning each
ics 57(2):450–456. treatment to several patients and assessing the outcome. Traditionally,
a single preliminary experiment would be conducted, after which the
apparently better-performing treatment would be adopted and the worse-
performing treatment discarded. However, thompson argued that one
should never eliminate either treatment at any stage of investigation,
even in the face of overwhelming evidence. Rather, he proposed a perpet-
ual clinical trial modeled as what we now call a two-armed bandit: the
possible treatments represent alternatives available to the clinician, and
patient outcomes determine the rewards. We can now consider assigning
treatments for a sequence of patients guided by our evolving knowledge,
hoping to efficiently and confidently determine the optimal treatment.
To conduct such a sequential clinical trial effectively, thompson
proposed maintaining a belief over which treatment is better in light
of available evidence and always selecting the nominally better treat-
ment according to its posterior probability of superiority. This is the
eponymous Thompson sampling policy (§ 7.9), which elegantly addresses

292
 12.5. multi-armed bandits to infinite-armed bandits

the exploration–exploitation dilemma. The more evidence we have in

favor of one treatment, the more often we prescribe it, exploiting the
apparently better choice. Meanwhile, we maintain a diminishing but
nonzero probability of selecting the other treatment, forcing continual
exploration until the better treatment becomes obvious. In the long term,
we will eventually assign the correct treatment to new patients with
probability approaching certainty.
thompson’s work was in retrospect groundbreaking, but its signifi-
cance was perhaps not fully realized at the time. However, Thompson
sampling for multi-armed bandits has recently enjoyed an explosion
of attention due to impressive empirical performance54, 55 and strong 54 o. chapelle and l. li (2011). An Empirical Eval-
theoretical guarantees (chapter 10).56, 57, 58 The first direct application of uation of Thompson Sampling. neurips 2011.
Thompson sampling to Bayesian optimization is due to shahriari et al.,59 55 o.-c. granmo (2010). Solving two-armed
who adopted an efficient approximation first proposed by hernández- Bernoulli bandit problems using a Bayesian
learning automaton. International Journal of
lobato et al. in the context of entropy search.60 Intelligent Computing and Cybernetics 3(2):
Although thompson had introduced bandits in the early 1930s, con- 207–234.
certed effort began with robbins’s landmark 1952 reintroduction and 56 s. agrawal and n. goyal (2012). Analysis of
analysis of the problem.61 This work introduced the modern formulation Thompson Sampling for the Multi-armed Ban-
of multi-armed bandits presented in § 7.7, where an arbitrary, unknown dit Problem. colt 2012.
reward distribution is associated with each arm and the agent seeks a pol- 57 d. russo and b. van roy (2014). Learning to
icy to maximize the expected cumulative reward. robbins explores this Optimize via Posterior Sampling. Mathematics
of Operations Research 39(4):1221–1243.
problem in the special case of a two-armed bandit with Bernoulli rewards,
demonstrating that a simple adaptive policy (“switch on lose, stay on 58 d. russo and b. van roy (2016). An
Information-Theoretic Analysis of Thompson
win”) can achieve better performance than nonadaptive or random poli- Sampling. Journal of Machine Learning Re-
cies. He also presents a family of policies that can achieve asymptotically search 17(68):1–30.
optimal behavior for any reward distributions. These policies are defined 59 b. shahriari et al. (2014). An Entropy Search
by a simple mechanism that explicitly forces continual but decreasingly Portfolio for Bayesian Optimization. arXiv:
frequent exploration of both arms according to a pregenerated schedule, 1406.4625 [stat.ML].
effectively – if crudely – balancing exploration and exploitation. 60 j. m. hernández-lobato et al. (2014). Predic-
The simple policies proposed by robbins eventually achieve near- tive Entropy Search for Efficient Global Opti-
mization of Black-box Functions. neurips 2014.
optimal cumulative reward, but they are not very efficient in the sense
61 h. robbins (1952). Some Aspects of the Se-
of achieving that behavior quickly. robbins returned to this problem
quential Design of Experiments. Bulletin of
three decades later to further address the issue of efficiency.62 lai and the American Mathematical Society 58(5):527–
robbins introduced policies that dynamically trade off exploration and 535.
exploitation by maximizing an upper confidence bound on the reward 62 t. l. lai and h. robbins (1985). Asymptotically
distributions of the arms (§ 7.8), and demonstrated that these policies Efficient Adaptive Allocation Rules. Advances
are both asymptotically optimal and efficient. auer et al. later proved in Applied Mathematics 6(1):4–22.
that bandit policies based on maximizing upper confidence bounds can
provide strong guarantees not only asymptotically but also in the finite-
budget case.63 Interestingly, kushner had proposed optimization policies 63 p. auer et al. (2002). Finite-time Analysis
based on maximizing an upper confidence bound of the objective func- of the Multiarmed Bandit Problem. Machine
Learning 47(2–3):235–256.
tion in the continuous case two decades earlier than lai and robbins’s
work,21 although he did not prove any performance guarantees for this
procedure and abandoned the idea in favor of maximizing probability of 64 d. d. cox and s. john (1992). A Statistical
improvement. This optimization policy would be rediscovered several Method for Global Optimization. smc 1992.
times, including by cox and john64 and by srinivas et al.,65 who estab- 65 n. srinivas et al. (2010). Gaussian Process Op-
lished theoretical guarantees on the rate of convergence of this policy in timization in the Bandit Setting: No Regret
the Gaussian process case, as discussed in § 10.4. and Experimental Design. icml 2010.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 293

 a brief history of bayesian optimization

12.6 what’s next?

The mathematical foundations of Bayesian optimization, as presented
in this book and chronicled in this chapter, are by now well established.
However, Bayesian optimization continues to deliver impressive perfor-
mance on a wide variety of problems, and the field continues to develop
at a rapid pace in light of this success. What big challenges remain on
the horizon? We speculate on some potential opportunities below.
Gaussian process modeling in high • Effective modeling of objective functions remains a challenge, especially
dimension: § 3.5, p. 61 in high dimension. Meanwhile, methods such as stochastic gradient de-
scent routinely (locally) optimize objectives in millions of dimensions,
and are “unreasonably effective” at doing so – all with very weak guid-
ance. Can we bridge this gap, either by extending or refining approaches
for Gaussian process modeling, or by exploring another model class?
moving beyond one-step lookahead: § 7.10, • Nonmyopic policies that reach beyond one-step lookahead have shown
p. 150 impressive empirical performance in initial investigations. However,
these policies have not yet been widely adopted, presumably due to their
(sometimes significantly) greater computational cost compared to myopic
alternatives. The continued development of efficient, yet nonmyopic
policies remains a promising avenue of research.
• A guiding philosophy in Bayesian optimization has been to “take the
human out of the loop” and hand over complete control of experimental
66 b. shahriari et al. (2016). Taking the Human design to an algorithm.66 This paradigm has demonstrated remarkable
Out of the Loop: A Review of Bayesian Op- success on “black box” problems, where the user has little understanding
timization. Proceedings of the ieee 104(1):148–
175.
of the system being optimized. However, in settings such as scientific
discovery, the user has a deep understanding of and intuition for the
This is an excellent companion survey on mechanisms driving the system of interest, and we should perhaps con-
Bayesian optimization for reader who have
reached this point but still want more. sider how to “bring them back into the loop.” One could imagine an
ecosystem of cooperative tools that enable Bayesian optimization algo-
rithms to benefit from user knowledge while facilitating the presentation
of experimental progress and evolving model beliefs back to the users.
• In the author’s experience, many consumers of Bayesian optimization
have experimental goals that are not perfectly captured by any of the
utility functions: § 6, p. 109 common utility functions used in Bayesian optimization – for example,
users may want to ensure adequate coverage of the domain or to find a
diverse set of locations with high objective values. Although the space
of Bayesian optimization policies is fairly crowded, there is still room
for innovation. If the ecosytem of cooperative tools envisioned above
were realized, we might consider tuning the utility function “on the fly”
via interactive preference elicitation.
• Bayesian optimization is not a particularly user-friendly approach, as
effective optimization requires careful modeling of the system of interest.
However, model construction remains something of a “black art,” even in
the machine learning community, and thus considerable machine learn-
ing expertise can be required to get the most out of this approach. Can
we build turnkey Bayesian optimization systems that achieve acceptable
performance even in the absence of clear prior beliefs?

294
 THE GAUSSIAN DISTRIBU TION

The Gaussian (or normal) distribution is a fundamental probability dis-

tribution in probabilistic modeling and inference. Gaussian distributions
are frequently encountered in Bayesian optimization as they serve as
the foundation of Gaussian processes, an infinite-dimensional extension
appropriate for reasoning about unknown objective functions. In this
chapter we provide a brief introduction to finite-dimensional Gaussian
a
Gaussian processes: chapters 2–4, p. 15
distributions and establish important properties referenced throughout
this book. We will begin with the univariate (one-dimensional) case, then
construct of the multivariate (vector-valued) case via linear transforma-
tions of univariate Gaussians.

a.1 univariate gaussian distribution

The univariate Gaussian distribution on a random variable 𝑥 ∈ ℝ has
two scalar parameters corresponding to its first two moments: 𝜇 = 𝔼[𝑥] mean, 𝜇
specifies the mean (also median and mode) and serves as a location
parameter, and 𝜎 2 = var[𝑥] specifies the variance and serves as a scale variance, 𝜎 2
parameter.

Probability density function and degenerate case

When the variance is nonzero, the distribution has the probability density
function


N (𝑥; 𝜇, 𝜎 2 > 0) = 𝑍 −1 exp − 21 𝑧 2 , (a.1)
√
where 𝑍 = 2𝜋𝜎 is a normalization constant and 𝑧 is the familiar 𝑧-score 𝑧
of 𝑥: −3 0 3
𝑥 −𝜇
𝑧= . (a.2) A univariate Gaussian probability density
𝜎 function N (𝑥; 𝜇, 𝜎 2 ) as a function of the
𝑧-score.
This pdf is illustrated in the margin. The probability density is rapidly
decreasing with the magnitude of the 𝑧-score, with for example 99.7% of
the density lying in the interval |𝑧| ≤ 3, or 𝑥 ∈ (𝜇 ± 3𝜎).
In the degenerate case 𝜎 2 = 0, the distribution collapses to a point degenerate case, 𝜎 2 = 0
mass at the mean and a probability density function does not exist. We
can express this case with the Dirac delta distribution:

N (𝑥; 𝜇, 𝜎 2 = 0) = 𝛿 (𝑥 − 𝜇). (a.3)

Standard normal distribution

The special case of zero mean and unit variance is called the standard
normal distribution and enjoys a privileged role. We notate its density
with the compact notation 𝜙 (𝑥) = N (𝑥; 0, 12 ). The cumulative density standard normal pdf, 𝜙
function of the standard normal cannot be expressed in terms of elemen-
tary functions, but is such an important quantity that it also merits its

This material will be published by Cambridge University Press as Bayesian Optimization. 295
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
the gaussian distribution

standard normal cdf, Φ own special notation:

∫𝑦
Φ(𝑦) = Pr(𝑥 < 𝑦) = 𝜙 (𝑥) d𝑥 .
−∞

expressing arbitrary pdfs and cdfs in terms We can write the pdf and cdf of an arbitrary nondegenerate Gausisan
of the standard normal distribution N (𝑥; 𝜇, 𝜎 2 ) in terms of the standard normal pdf and cdf by
appropriately rescaling and translating arguments to their 𝑧-scores (a.1):
1 𝑥 − 𝜇  𝑦 − 𝜇 
𝑝 (𝑥) = 𝜙 ; Pr(𝑥 < 𝑦) = Φ ,
𝜎 𝜎 𝜎
where the mulitplicative factor in the pdf guarantees normalization.

Affine transformations
The family of univariate Gaussian distributions is closed under affine
transformations, which simply translate and rescale the distribution and
adjust its moments accordingly. If 𝑥 has distribution N (𝑥; 𝜇, 𝜎 2 ), then
the transformation 𝜉 = 𝑎𝑥 + 𝑏 has distribution

𝑝 (𝜉) = N (𝜉; 𝑎𝜇 + 𝑏, 𝑎 2𝜎 2 ).

The above results allow us to interpret any univariate Gaussian distribu-

tion as a translated and scaled copy of the standard normal after applying
the transformation 𝑥 ↦→ 𝜇 + 𝜎𝑥. This process is illustrated in the margin,
where a standard
√ normal distribution is transformed via the mapping
𝑥 ↦→ 1 + 𝑥/ 2, resulting in a new Gaussian with increased mean 𝜇 = 1
A standard normal pdf (dashed) and the
pdf after √applying the transformation and decreased variance 𝜎 2 = 1/2. The pdf is appropriately translated and
𝑥 ↦→ 1 + 𝑥/ 2 (solid). rescaled.

a.2 multivariate gaussian distribution

The multivariate Gaussian distribution extends the univariate case to an
arbitrary random vector x ∈ ℝ𝑑. We will provide an explicit construction
of the multivariate Gaussian distribution as the result of applying an
affine transformation to independent univariate standard normal random
variables, extending the properties noted at the end of the previous
section.

Standard multivariate normal and construction of general case

standard multivariate normal First we construct the standard multivariate normal distribution, rep-
resented by a random vector z ∈ ℝ𝑑 whose entries are independent
Î
standard univariate normal random variables: 𝑝 (z) = 𝑖 𝜙 (𝑧𝑖 ). It is clear
from construction that the mean of this distribution is the zero vector
and its covariance is the identity matrix:

𝔼[z] = 0; cov[z] = I, (a.4)

and we will denote its density with 𝑝 (z) = N (z; 0, I).

296
 a.2. multivariate gaussian distribution

As before, this will serve as the basis of the general multivariate constructing general case via affine
case by considering arbitrary affine transformations of this “standard” transformations of standard normal vectors
example. Suppose x ∈ ℝ𝑑 is a vector-valued random variable and z ∈ ℝ,𝑘
𝑘 ≤ 𝑑, is a 𝑘-dimensional standard multivariate normal vector. If we can
write
x = 𝝁 + 𝚲z (a.5)
for some vector 𝝁 ∈ ℝ𝑑 and 𝑑 × 𝑘 matrix 𝚲, then x has a multivariate
normal distribution. We can compute its mean and covariance directly
from (a.4–a.5):

𝔼[x] = 𝝁; cov[x] = 𝚲𝚲> = 𝚺.

This property completely characterizes the distribution. As in the parameterization in terms of first two
univariate case, we can interpret every multivariate normal distribution moments
as an affine transformation of a (possibly lower-dimensional) standard
normal vector. We again parameterize this family by its first two mo-
ments: the mean vector 𝝁 and the covariance matrix 𝚺. This covariance mean vector and covariance matrix, (𝝁, 𝚺)
matrix is necessarily symmetric and positive semidefinite, which means positive semidefinite
all its eigenvalues are nonnegative. We can factor any such matrix as
𝚺 = 𝚲𝚲>, allowing us to recover the underlying transformation (a.5),
although 𝚲 need not be unique.

Probability density function and degenerate case

If 𝚲 has full rank 𝑑, then the range of (a.5) is all of ℝ𝑑 and a probability
density function exists. This condition further implies that the covariance
matrix 𝚺 is positive definite; that is, its eigenvalues are strictly positive, positive definite
implying its determinant is positive and the matrix is invertible. The
distribution has a probability density function in this case analogous to
the univariate pdf (a.1):


N (x; 𝝁, 𝚺) = 𝑍 −1 exp − 21 Δ2 . (a.6)

Here 𝑍 again represents a normalization constant: normalization constant, 𝑍

√︁ 𝑑 1
𝑍= |2𝜋 𝚺| = (2𝜋) 2 |𝚺| 2 , (a.7)

and Δ represents the Mahalanobis distance, a multivariate analog of the Mahalanobis distance, Δ
(absolute) 𝑧-score (a.2):

Δ2 = (x − 𝝁) > 𝚺−1 (x − 𝝁). (a.8)

It is easy to verify that these definitions are compatible with the univari- compatibility with univariate case
ate case when 𝑑 = 1. Note in that case the condition of 𝚺 being positive
definite reduces to the previous condition for nondegeneracy, 𝜎 2 > 0.
The dependence of the multivariate Gaussian density on x is entirely
through the value of the Mahalanobis distance Δ. To gain some geometric
insight into the probability density, we can set this value to a constant and
compute isoprobability contours. In the case of the standard multivariate

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 297

 the gaussian distribution

Figure a.1: Isoprobability contours Δ = 1 𝑥2 𝑥2 𝑥2

for the standard bivariate normal
distribution (left), a distribution
with diagonal covariance, scal- 𝑥1 𝑥1 𝑥1
ing the probability along the axes
(middle), and a distribution with
nonzero off-diagonal covariance,
tilting the probability (right). The
standard normal contour is shown    
for reference on the latter two ex- 0.5 0 1 −0.9
𝚺=I 𝚺= 𝚺=
amples. 0 2 −0.9 1

Gaussian distribution, the Mahalanobis distance reduces to the normal

Euclidean distance, and the set of points satisfying Δ = 𝑐 > 0 is then
a sphere of radius 𝑐 centered at the origin – see the illustration in the
margin.
We can now understand the geometry of the general case N (x; 𝝁, 𝚺)
purely in terms of the affine transformation (a.5) translating and warp-
ing the spherical distribution of the standard normal. Taking this view,
the action of multiplying by 𝚲 warps the isoprobability contours into
Probability density and circular isoprobabil- ellipsoids, which are then translated to the new center 𝝁. The standard
ity contours of a standard bivariate normal
distribution.
theory of linear maps then gives further insight: the principal axes of the
ellipsoids are given by the eigenvectors of 𝚲, and their axis semilengths
are given by the eigenvalues scaled by 𝑐. See figure a.1 for an illustration.
degenerate case, |𝚺 | = 0 The probability density function does not exist when 𝚲 (and thus
𝚺) is rank-deficient, as the range of x would then be restricted to the
lower-dimensional affine subspace {𝝁 + 𝚲z | z ∈ ℝ𝑑 } (a.5). However, it
1 On this space, the pdf is similar to (a.6–a.8) is still possible to define a probability density function in this degenerate
but replaces the determinant with the pseu- case by restricting the support to this subspace.1 This is analogous to the
dodeterminant and the inverse with the pseu-
doinverse. If 𝚺 = 0, the distribution is a Dirac
degenerate univariate case (a.3), where probability was restricted to the
delta on 𝝁. zero-dimensional subspace containing the mean only, {𝜇}.

Affine transformations
The multivariate Gaussian distribution has a number of convenient
mathematical properties, many of which follow immediately from the
characterization in (a.5). First it is obvious that any affine transfor-
mation of a multivariate normal distributed vector is also multivari-
ate normal, as affine transformations are closed under composition. If
distribution of affine transformation Ax + b 𝑝 (x) = N (x; 𝝁, 𝚺), then 𝝃 = Ax + b has distribution

𝑝 (𝝃 ) = N (𝝃 ; A𝝁 + b, A𝚺A> ). (a.9)

Further, if we apply this result with the transformation

     
I 0 x
x ↦→ x+ = ,
A b 𝝃

298
 a.2. multivariate gaussian distribution

we can see that x and 𝝃 in fact have a joint Gaussian distribution: joint distribution with affine transformations
     !
x 𝝁 𝚺A>
; (a.10)
𝚺
𝑝 (x, 𝝃 ) = N , .
𝝃 A𝝁 + b A𝚺 A𝚺A>

Sampling
The characterization of the multivariate normal in terms of affine trans-
formations of standard normal random variables (a.5) also suggests a
simple algorithm for drawing samples from the distribution. Given an
arbitrary multivariate normal distribution N (x; 𝝁, 𝚺), we first factor
the covariance as 𝚺 = 𝚲𝚲>, where 𝚲 has size 𝑑 × 𝑘; when 𝚺 is positive
definite, the Cholesky decomposition is the canonical choice. We now
sample a 𝑘-dimensional standard normal vector z by sampling each entry
independently from a univariate standard normal; routines for this task
are readily available. Finally, we transform this vector appropriately to
provide a sample of x: z ↦→ 𝝁 + 𝚲z = x. This procedure entails one-time
O(𝑑 3 ) work to compute 𝚲 (which can be reused), followed by O(𝑑 2 )
work to produce each sample.

Marginalization
Often we will have a vector x with a multivariate Gaussian distribution, 2 We can first permute x if required; as a linear
but only be interested in reasoning about a subset of its entries. Suppose transformation, this will simply permute the
entries of 𝝁 and the rows/columns of 𝚺.
𝑝 (x) = N (x; 𝝁, 𝚺), and partition the vector into two components:2
 
x 𝑥2
x= 1 . (a.11)
x2
We partition the mean vector and covariance matrix in the same way:
     !
x1 𝝁1 𝚺11 𝚺12
𝑝 (x) = N ; , . (a.12)
x2 𝝁2 𝚺21 𝚺22

Now writing the subvector x1 as x1 = [I, 0] x and applying the affine

property (a.9), we have: 𝑥1

𝑝 (x1 ) = N (x1 ; 𝝁 1, 𝚺11 ). (a.13)

That is, to derive the marginal distribution of x1 we simply pick out the
corresponding entries of 𝝁 and 𝚺.

𝑥1
Conditioning
A bivariate Gaussian pdf 𝑝 (𝑥 1 , 𝑥 2 ) (top) and
Multivariate Gaussian distributions are also closed under conditioning the Gaussian marginal 𝑝 (𝑥 1 ) (bottom) (a.13).
on the values of given entries. Suppose again that 𝑝 (x) = N (x; 𝝁, 𝚺) and
partition x, 𝝁, and 𝚺 as before (a.11–a.12). Suppose now that we learn the
exact value of the subvector x2 . The posterior on the remaining entries
𝑝 (x1 | x2 ) remains Gaussian, with distribution
𝑝 (x1 | x2 ) = N (x1 ; 𝝁 1|2, 𝚺11 |2 ).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 299

the gaussian distribution

𝑥2 The posterior mean and covariance take the form of updates to the prior
moments in light of the revealed information:

𝝁 1 |2 = 𝝁 1 + 𝚺12 𝚺−1
22 (x2 − 𝝁 2 ); 𝚺11 |2 = 𝚺11 − 𝚺12 𝚺−1
22 𝚺21 . (a.14)

The mean is adjusted by an amount dependent on

1. the covariance between x1 and x2 , 𝚺12 ,
2. the uncertainty in x2 , 𝚺22 , and
𝑥1 3. the deviation of the observed values from the prior mean, (x2 − 𝝁 2 ).
Similarly, the uncertainty in x1 , 𝚺11 , is reduced by an amount dependent
on factors 1–2. Notably, the correction to the covariance matrix does not
depend on the observed values. Note that if x1 and x2 are independent,
then 𝚺12 = 0, and conditioning does not alter the distribution of x1 .

Sums of normal vectors

𝑥1
Suppose x and y are 𝑑-dimensional random vectors with joint multivari-
ate normal distribution      !
The pdfs of a bivariate Gaussian 𝑝 (𝑥 1 , 𝑥 2 ) x 𝝁 𝚺 P
(top) and the conditional distribution 𝑝 (x, y) = N ; , .
y 𝝂 P T
𝑝 (𝑥 1 | 𝑥 2 ) given the value of 𝑥 2 marked
by the dotted line (bottom) (a.14). The prior
marginal distribution 𝑝 (𝑥 1 ) is shown for Then recognizing their sum z = x + y = [I, I] [x, y] > as a linear transfor-
reference; the observation decreased both the mation and applying (a.9), we have:
mean and standard deviation.
𝑝 (z) = N (z; 𝝁 + 𝝂, 𝚺 + 2P + T).

independent case When x and y are independent, P = 0, and this simplifies to

𝑝 (z) = N (z; 𝝁 + 𝝂, 𝚺 + T), (a.15)

where the moments simply add.

Differential entropy
The differential entropy of a multivariate normal random variable x with
distribution 𝑝 (x) = N (x; 𝝁, 𝚺), expressed in nats, is
1
𝐻 [x] = 2 log |2𝜋𝑒𝚺|. (a.16)

In the univariate case 𝑝 (𝑥) = N (𝑥; 𝜇, 𝜎 2 ), this reduces to

𝐻 [𝑥] = 1
2 log 2𝜋𝑒𝜎 2. (a.17)

Sequences of normal random variables

If {x𝑖 } is a sequence of normal random variables with means {𝝁𝑖 } and
covariances {𝚺𝑖 } converging respectively to finite limits 𝝁𝑖 → 𝝁 and
𝚺𝑖 → 𝚺, then the sequence converges in distribution to a normal random
variable x with mean 𝝁 and covariance 𝚺.

300
 METHODS FOR APPROXIMATE BAYESIAN INFERENCE

In Bayesian optimization we occasionally face intractable posterior distri-

butions that must be approximated before we can proceed. The Laplace
approximation and Gaussian expectation propagation are two workhorses
of approximate Bayesian inference, and at least one will suffice in most
scenarios. Both result in Gaussian approximations to the posterior, espe-
cially convenient when working with Gaussian processes.
b
Approximate inference for gps: § 2.8, p. 39

b.1 the laplace approximation

Consider a vector-valued random variable x ∈ ℝ𝑑 with arbitrary prior
distribution 𝑝 (x). Suppose we obtain information D, yielding an in-
tractable posterior

𝑝 (x | D) = 𝑍 −1 𝑝 (x) 𝑝 (D | x)

that we wish to approximate. The Laplace approximation is based on

approximating the logarithm of the unnormalized posterior: unnormalized log posterior, Ψ

Ψ(x) = log 𝑝 (x) + log 𝑝 (D | x)

with a Taylor expansion around its maximum: maximum a posteriori point, x̂

x̂ = arg max Ψ(x).

x

Taking a second-order Taylor expansion around this point yields

Ψ(x) ≈ Ψ( x̂) − 21 (x − x̂) > H(x − x̂),

where H is the Hessian of the negative log posterior evaluated at x̂: Hessian of negative log posterior, H

𝜕2 Ψ
H=− ( x̂).
𝜕x 𝜕x>
Note the first-order term vanishes as we are expanding around a maxi-
mum. Exponentiating, we derive an approximation to the unnormalized
posterior:


𝑝 (x | D) ∝ exp Ψ(x) ≈ exp Ψ( x̂) exp − 21 (x − x̂) > H(x − x̂) .

We recognize this as proportional to a Gaussian distribution, yielding a

normal approximate posterior: Laplace approximation to posterior

𝑝 (x | D) ≈ 𝑞(x) = N (x; x̂, H−1 ). (b.1)

Through some accounting when normalizing (b.1), the Laplace approxi- Laplace approximation to normalizing
mation also gives an approximation to the normalizing constant 𝑍 : constant, 𝑍ˆla

𝑍 ≈ 𝑍ˆla = (2𝜋) 2 |H| − 2 exp Ψ( x̂).

𝑑 1
(b.2)

This material will be published by Cambridge University Press as Bayesian Optimization. 301
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 methods for approximate bayesian inference

true distribution Laplace approximation

unnormalized Laplace approximation

Figure b.1: A Laplace approximation to a one-dimensional posterior distribution. The left panel shows the Laplace
approximation before normalization, and the right panel afterwards.

The Laplace approximation procedure is illustrated in figure b.1,

where we show the approximate posterior both before and after nor-
malization. The posterior density is an excellent local approximation
around the maximum but is not a great global fit as a significant fraction
of the true posterior mass is ignored. However, the Laplace approxima-
tion is remarkably simple and general and is sometimes the only viable
approximation scheme.

b.2 gaussian expectation propagation

Expectation propagation (ep) is a technique for approximate Bayesian
inference that enjoys some use in Bayesian optimization. We will give
a brief and incomplete introduction that should nonetheless suffice for
1 t. p. minka (2001). A family of algorithms for common applications in this context. A complete introduction can be
approximate Bayesian inference. Ph.D. thesis. found in minka’s thesis,1 and cunningham et al. provide in-depth advice
Massachusetts Institute of Technology.
regarding efficient and stable computation for the rank-one case we
2 j. p. cunningham et al. (2011). Gaussian Prob- consider here.2
abilities and Expectation Propagation. arXiv:
1111.6832 [stat.ML]
Consider a multivariate Gaussian random variable 𝝃 with distribution

𝑝 (𝝃 ) = N (𝝃 ; 𝝁 0, 𝚺0 ).

3 These scalar products often simply pick out Suppose we obtain information D about 𝝃 in the form of a collection of
single entries of 𝝃 , in which case we can factors, each of which specifies the likelihood of a scalar product 𝑥 = a>𝝃
slightly simplify notation. However, we some-
times require this more general formulation. associated with that factor.3 We consider the posterior distribution
Ö
𝑝 (𝝃 | D) = 𝑍 −1 𝑝 (𝝃 ) 𝑡𝑖 (𝑥𝑖 ), (b.3)
𝑖

where 𝑖th factor 𝑡𝑖 informs our belief about 𝑥𝑖 = a𝑖>𝝃 . Unfortunately,

this posterior is intractable except the notable case when all factors are
Gaussian.
Gaussian expectation propagation proceeds by replacing each of the
factors with an unnormalized Gaussian distribution:

𝑡𝑖 (𝑥𝑖 ) ≈ 𝑡˜𝑖 (𝑥𝑖 ) = 𝑍˜𝑖 N (𝑥𝑖 ; 𝜇˜𝑖 , 𝜎˜ 𝑖2 ).

302
 b.2. gaussian expectation propagation

Here (𝑍˜𝑖 , 𝜇˜𝑖 , 𝜎˜ 𝑖2 ) are called site parameters for the approximate factor 𝑡˜𝑖 , ˜ 𝜇,
site parameters, (𝑍, ˜ 𝜎˜ 2 )
which we may design to optimize the fit. We will consider this issue fur-
ther shortly. Given arbitrary site parameters for each factor, the resulting Gaussian ep approximate posterior, 𝑞 (𝝃 )
approximation to (b.3) is
Ö
𝑝 (𝝃 | D) ≈ 𝑞(𝝃 ) = 𝑍ˆep
−1
𝑝 (𝝃 ) 𝑡˜𝑖 (𝑥𝑖 ). (b.4)
𝑖

As a product of Gaussians, the approximate posterior is also Gaussian: 4 The updated covariance incorporates only a
series of rank-one updates, which can be ap-
𝑞(𝝃 ) = N (𝝃 ; 𝝁, 𝚺), (b.5) plied using the Sherman–Morrison formula.

with parameters:4 Gaussian ep parameters, (𝝁, 𝚺)

 ∑︁ 𝜇˜   ∑︁ 1  −1
𝝁 = 𝚺 𝚺−1
0 𝝁0 +
𝑖
a ;
2 𝑖
𝚺 = 𝚺−1
0 + a a >
.
𝑖
˜
𝜎 𝑖 𝑖
𝜎˜ 𝑖2 𝑖 𝑖

Gaussian ep also yields an approximation of the normalizing constant 𝑍 , Gaussian ep approximation to normalizing
if desired: constant, 𝑍ˆep

N (0; 𝝁, 𝚺) Ö ˜
𝑍 ≈ 𝑍ˆep = 𝑍 N (0; 𝜇˜𝑖 , 𝜎˜ 𝑖2 ). (b.6)
N (0; 𝝁 0, 𝚺0 ) 𝑖 𝑖

What remains to be determined is an effective means of choosing the setting site parameters
site parameters to maximize the approximation fidelity. One reasonable
goal would be to minimize the Kullback–Leibler (kl) divergence between
the true and approximate distributions; for our Gaussian approximation
(b.5), this is achieved through moment matching.2 Unfortunately, de-
termining the moments of the true posterior (b.3) may be difficult, so
expectation propagation instead matches the marginal moments for expectation propagation approximately
each of the {𝑥𝑖 }, approximately minimizing the kl divergence. This is ac- minimizes kl divergence
complished through an iterative procedure where we repeatedly sweep
over each of the approximate factors and refine its parameters until
convergence.
We initialize all site parameters to (𝑍, ˜ 𝜇,
˜ 𝜎˜ 2 ) = (1, 0, ∞); with these site parameter initialization
choices the approximate factors drop away, and our initial approxima-
tion is simply the prior: (𝝁, 𝚺) = (𝝁 0, 𝚺0 ). Now we perform a series of
updates for each of the approximate factors in turn. These updates take a
convenient general form, and we will drop factor index subscripts below
to simplify notation.
Let 𝑡˜ (𝑥) = 𝑍˜ N (𝑥; 𝜇,
˜ 𝜎˜ 2 ) be an arbitrary factor in our approximation
(b.4). The idea behind expectation propagation is to drop this factor from
the approximation entirely, forming the cavity distribution: cavity distribution, 𝑞¯

𝑞(𝝃 )
¯ )=
𝑞(𝝃 ,
𝑡˜ (𝑥)

and replace it with the true factor 𝑡 (𝑥), forming the tilted distribution tilted distribution
¯ ) 𝑡 (𝑥). The tilted distribution is closer to the true posterior (b.3) as
𝑞(𝝃
the factor in question is no longer approximated. We now adjust the site

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 303

 methods for approximate bayesian inference

Gaussian ep approximation
Figure b.2: A Gaussian ep approximation to the distribution
in figure b.1.

parameters to minimize the kl divergence between the tilted distribution

and the new approximation 𝑞 0(𝝃 ) = 𝑞(𝝃 ¯ ) 𝑡˜0(𝑥):
 
˜ 𝜇,
(𝑍, ˜ 𝜎˜ 2 ) = arg min 𝐷 kl 𝑞(𝝃
¯ ) 𝑡 (𝑥) k 𝑞 0(𝝃 )

by matching zeroth, first, and second moments.

Because Gaussian distributions are closed under marginalization, we
5 m. seeger (2008). Expectation Propagation for can simplify this procedure by manipulating only marginal distributions
Exponential Families. Technical report. Univer- for 𝑥 rather than the full joint distribution.5 The marginal belief about
sity of California, Berkeley.
𝑥 = a>𝝃 in our current approximation (b.5) is:

𝑞(𝑥) = N (𝑥; 𝜇, 𝜎 2 ); 𝜇 = a𝑖>𝝁; 𝜎 2 = a𝑖> 𝚺a𝑖 .

By dividing by the approximate factor 𝑡˜ (𝜉), we arrive at the marginal

cavity distribution, which is Gaussian:

¯
𝑞(𝑥) ¯ 𝜎¯ 2 );
= N (𝑥; 𝜇, 𝜇¯ = 𝜎¯ 2 (𝜇𝜎 −2 − 𝜇˜𝜎˜ −2 ); 𝜎¯ 2 = (𝜎 −2 − 𝜎˜ −2 ) −1
. (b.7)

Consider the zeroth moment of the marginal tilted distribution:

∫
¯ 𝜎¯ 2 ) d𝑥;
𝑍 = 𝑡 (𝑥) N (𝑥; 𝜇, (b.8)

¯ 𝜎¯ 2 ). If we
this quantity clearly depends on the cavity parameters ( 𝜇,
define
𝜕 log 𝑍 𝜕 log 𝑍
𝛼= ; 𝛽= ; (b.9)
𝜕 𝜇¯ 𝜕𝜎¯ 2
6 t. minka (2008). ep: A quick reference. and an auxiliary variable 𝛾 = (𝛼 2 − 2𝛽) −1
, then we may achieve the
desired moment matching by updating the site parameters to:6
√︁ √︁

𝜇˜ = 𝜇¯ +𝛼𝛾; 𝜎˜ 2 = 𝛾 − 𝜎¯ 2 ; 𝑍˜ = 𝑍 2𝜋 𝜎˜ 2 + 𝜎¯ 2 exp 21 𝛼 2𝛾 . (b.10)

This completes our update for the chosen factor; the full ep procedure
repeatedly updates each factor in this manner until convergence. The
result of Gaussian expectation propagation for the distribution from
figure b.1 is shown in figure b.2. The fit is good and reflects the more
global nature of the expectation propagation scheme achieved through
moment matching rather than merely maximizing the posterior.
A convenient aspect expectation propagation is that incorporating
a new factor only requires computing the zeroth moment against an
arbitrary normal distribution (b.8) and the partial derivatives in (b.9). We
provide these computations for several useful factor types below.

304
 b.2. gaussian expectation propagation

factor, 𝑡 (𝑥) Gaussian ep approximation

tilted distribution

𝑎 𝑎

Figure b.3: A Gaussian ep approximation to a one-dimensional normal distribution truncated at 𝑎.

Figure b.4: A Gaussian ep approximation to a joint nor-

mal distribution conditioned on one coordi-
nate being less than the other, correspond-
ing to the dashed boundary. Contours of
the tilted distribution are shown in the left
panel, and contours of the approximate pos-
terior in the right panel.

Truncating a variable
A common use of expectation propagation in Bayesian optimization
is to approximately constrain a Gaussian random variable 𝑥 to be less
than a threshold 𝑎. We may capture this information by a single factor
𝑡 (𝑥) = [𝑥 < 𝑎]. In the context of expectation propagation, we must
consider the normalizing constant of the tilted distribution, which is a
truncated normal:
∫
𝑎 − 𝜇¯
¯ 𝜎¯ 2 ) d𝑥 = Φ(𝑧);
𝑍 = [𝑥 < 𝑎] N (𝑥; 𝜇, 𝑧= . (b.11)
𝜎¯
The required quantities for an expectation propagation update are now
(b.10):
  −1
𝜙 (𝑧) 𝑧𝛼 𝜎¯ 𝜙 (𝑧)
𝛼 =− ; 𝛽= ; 𝛾 =− +𝑧 . (b.12)
Φ(𝑧)𝜎¯ 2𝜎¯ 𝛼 Φ(𝑧)
A Gaussian ep approximation to a truncated normal distribution is il-
lustrated in figure (b.3). The fit is good, but not perfect: approximately
5% of its mass exceeds the threshold. This inaccuracy is the price of
approximation.
We may also apply this approach to approximately condition our conditioning on one variable being greater
belief on 𝝃 on one entry being dominated by another: 𝜉𝑖 < 𝜉 𝑗 . Consider than another
the vector a where 𝑎𝑖 = 1, 𝑎 𝑗 = −1 and all other entries are zero. Then
𝑥 = a>𝝃 = 𝜉𝑖 − 𝜉 𝑗 . The condition is now equivalent to [𝑥 < 0], and we
can proceed as outlined above. This approximation is illustrated for a
bivariate normal in figure b.4; again the fit appears reasonable.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 305

methods for approximate bayesian inference

factor, 𝑡 (𝑥) Gaussian ep approximation

tilted distribution

95% ci for 𝑎 95% ci for 𝑎

Figure b.5: A Gaussian ep approximation to a one-dimensional normal distribution truncated at an unknown threshold 𝑎
with the marked 95% credible interval.

Truncation at an uncertain threshold

A sometimes useful extension of the above is to consider truncation at
an uncertain threshold 𝑎. Suppose we have a Gaussian belief about 𝑎:

𝑝 (𝑎) = N (𝑎; 𝜇, 𝜎 2 ).

Integrating the hard truncation factor [𝑥 < 𝑎] against this belief yields
the following “soft truncation” factor:
∫ 𝜇 −𝑥 
𝑡 (𝑥) = [𝑥 < 𝑎] 𝑝 (𝑎) d𝑎 = Φ . (b.13)
𝜎
We consider again the normalizing constant of the tilted distribution:
∫ 
𝜇 −𝑥 𝜇 − 𝜇¯
𝑍= Φ ¯ 𝜎¯ 2 ) d𝑥 = Φ(𝑧);
N (𝑥; 𝜇, 𝑧=√ ,
𝜎 𝜎 2 + 𝜎¯ 2
√
Defining 𝑠 = 𝜎 2 + 𝜎¯ 2 , we may compute:
  −1
𝜙 (𝑧) 𝑧𝛼 𝑠 𝜙 (𝑧)
𝛼 =− ; 𝛽= ; 𝛾 =− +𝑧 .
Φ(𝑧)𝑠 2𝑠 𝛼 Φ(𝑧)

The hard truncation formulas above may be interpreted as a special case

of this result by setting (𝜇, 𝜎 2 ) = (𝑎, 0). This procedure is illustrated in
figure b.5, where we softly truncate a one-dimensional Gaussian distri-
bution.

306
 GRADIEN TS

Under mild continuity assumptions, for Gaussian processes conditioned

with exact inference, we may compute the gradient of both the log
marginal likelihood with respect to model hyperparameters and of the
posterior predictive moments with respect to observation location. The
former aids in maximizing or sampling from the model posterior, and the
latter in maximizing acquisition functions derived from the predictive
c
distribution. In modern software these gradients are often computed via
automatic differentiation, but we present their functional forms here to
offer insight into their behavior.
We will consider a function 𝑓 : ℝ𝑑 → ℝ with distribution GP (𝑓 ; 𝑚, 𝐾),
observed with independent (but possibly hetereoskedastic) additive Gaus-
sian noise:
𝑝 (𝑦 | 𝑥, 𝜙, 𝜎𝑛 ) = N (𝑦; 𝜙, 𝜎𝑛2 ),
where the noise scale 𝜎𝑛 may optionally depend on 𝑥. We will notate the
prior moment and noise scale functions with:
𝑚(𝑥; 𝜽 ); 𝐾 (𝑥, 𝑥 0; 𝜽 ); 𝜎𝑛 (𝑥; 𝜽 ),
and assume these are differentiable with respect to observation location
and any parameters they may have. In a slight abuse of notation, we will
use 𝜽 to indicate a vector collating the values of all hyperparameters of
this model.
We will consider an arbitrary set of observations D = (x, y). We will
write the prior moments of 𝝓 = 𝑓 (x) and the noise covariance associated
with these observations as:
𝝁 = 𝑚(x; 𝜽 ); 𝚺 = 𝐾 (x, x; 𝜽 ); N = diag 𝜎𝑛2 (x; 𝜽 ),
all of which have implicit dependence on the hyperparameters. It will
also be useful to introduce notation for two repeating quantities:
V = cov[y | x, 𝜽 ] = 𝚺 + N; 𝜶 = V−1 (y − 𝝁).

c.1 gradient of log marginal likelihood

The log marginal likelihood of the data is (4.8):
 
L(𝜽 ) = log 𝑝 (y | x, 𝜽 ) = − 12 𝜶 > (y − 𝝁) + log |V| + 𝑛 log 2𝜋 .
The partial derivatives with respect to mean function parameters have
the form:
𝜕L 𝜕𝝁
= 𝜶> ,
𝜕𝜃 𝜕𝜃
and partial derivatives with respect to covariance function and likelihood
parameters (that is, the parameters of V) take the form:
"  #
𝜕L 1 > 𝜕V −1 𝜕V
= 𝜶 𝜶 − tr V .
𝜕𝜃 2 𝜕𝜃 𝜕𝜃

This material will be published by Cambridge University Press as Bayesian Optimization. 307
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 gradients

c.2 gradient of predictive distribution with respect to location

For a given observation location 𝑥, let us define the vectors

k = 𝐾 (x, 𝑥); 𝜷 = V−1 k.

The posterior moments of 𝜙 = 𝑓 (𝑥) are (2.19):

𝜇 = 𝑚(𝑥) + 𝜶 > k; 𝜎 2 = 𝐾 (𝑥, 𝑥) − 𝜷 > k,

and the partial derivatives of these moments with respect to observation

location are:
𝜕𝜇 𝜕𝑚 𝜕k 𝜕𝜎 2 𝜕𝐾 (𝑥, 𝑥) 𝜕k
= − 𝜶> ; = − 2𝜷 > . (c.1)
𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥
The predictive distribution for a noisy observation 𝑦 at 𝑥 is (8.5):

𝑝 (𝑦 | 𝑥, D, 𝜎𝑛2 ) = N (𝑦; 𝜇, 𝑠 2 ); 𝑠 2 = 𝜎 2 + 𝜎𝑛2 ,

and the partial derivative of the predictive variance and standard devia-
tion with respect to 𝑥 are:

𝜕𝑠 2 𝜕𝜎 2 𝜕𝜎𝑛2 𝜕𝑠 1 𝜕𝑠 2
= + ; = .
𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥 2𝑠 𝜕𝑥

c.3 gradients of common acqisition functions

Gradient of noisy expected improvement
this picks up the discussion from § 8.2, p. 163 To aid in the optimization of expected improvement, we may also com-
pute the gradient of 𝑔,1 and thus the gradient of expected improvement
1 w. scott et al. (2011). The Correlated Knowl- with respect to the proposed observation location 𝑥. This will require
edge Gradient for Simulation Optimization of several applications of the chain rule due to numerous dependencies
Continuous Parameters Using Gaussian Pro-
cess Regression. siam Journal on Optimization among the variables involved.
21(3):996–1026. First we note that the {𝑐𝑖 } values defining the intervals of dominance
(8.15) depend on the vectors a and b. In particular, for 2 ≤ 𝑖 ≤ 𝑛, 𝑐𝑖 is the
gradient of endpoints with respect to a, b 𝑧-value where the (𝑖 − 1)th and 𝑖th lines intersect, which occurs at
𝑎𝑖 − 𝑎𝑖−1
𝑐𝑖 = ,
𝑏𝑖−1 − 𝑏𝑖
making the dependence explicit. For 2 ≤ 𝑖 ≤ 𝑛, we have:
𝜕𝑐𝑖 1 𝜕𝑐𝑖 𝜕𝑐𝑖 𝜕𝑐𝑖 𝑎𝑖 − 𝑎𝑖−1 𝜕𝑐𝑖 𝜕𝑐𝑖
= ; =− = ; =− ,
𝜕𝑎𝑖 𝑏𝑖−1 − 𝑏𝑖 𝜕𝑎𝑖−1 𝜕𝑎𝑖 𝜕𝑏𝑖 (𝑏𝑖 − 𝑏𝑖−1 ) 2 𝜕𝑏𝑖−1 𝜕𝑏𝑖
and at the fixed endpoints at infinity 𝑖 ∈ {1, 𝑛 + 1}, we have
𝜕𝑐𝑖 𝜕𝑐𝑖
= = 0>
.
𝜕a 𝜕b
gradient of 𝑔 with respect to a and b Now we may compute the gradient of 𝑔(a, b) with respect to its inputs,
accounting for the implicit dependence of interval endpoints on these

308
 c.3. gradients of common acqisition functions

values:
𝜕𝑔  
= Φ(𝑐𝑖+1 ) − Φ(𝑐𝑖 )
𝜕𝑎𝑖
𝜕𝑐𝑖+1  
+ 𝑎𝑖 + 𝑏𝑖 𝑐𝑖+1 − [𝑖 ≤ 𝑛] (𝑎𝑖+1 + 𝑏𝑖+1𝑐𝑖+1 ) 𝜙 (𝑐𝑖+1 )
𝜕𝑎𝑖
𝜕𝑐𝑖  
− 𝑎𝑖 + 𝑏𝑖 𝑐𝑖 − [𝑖 > 1] (𝑎𝑖−1 + 𝑏𝑖−1𝑐𝑖 ) 𝜙 (𝑐𝑖 );
𝜕𝑎𝑖
𝜕𝑔  
= 𝜙 (𝑐𝑖 ) − 𝜙 (𝑐𝑖+1 )
𝜕𝑏𝑖
𝜕𝑐𝑖+1  
+ 𝑎𝑖 + 𝑏𝑖 𝑐𝑖+1 − [𝑖 ≤ 𝑛] (𝑎𝑖+1 + 𝑏𝑖+1𝑐𝑖+1 ) 𝜙 (𝑐𝑖+1 )
𝜕𝑏𝑖
𝜕𝑐𝑖  
− 𝑎𝑖 + 𝑏𝑖 𝑐𝑖 − [𝑖 > 1] (𝑎𝑖−1 + 𝑏𝑖−1𝑐𝑖 ) 𝜙 (𝑐𝑖 ).
𝜕𝑏𝑖
Here [𝑖 > 1] and [𝑖 ≤ 𝑛] represent the Iverson bracket. We will also gradient of a and b with respect to 𝑥
require the gradient of the a and b vectors with respect to 𝑥:
   
𝜕a 0 𝜕b 1 𝜕𝐾D (x,0 𝑥) 𝜕𝑠
= 𝜕𝜇 ; = −b . (c.2)
𝜕𝑥 𝜕𝑥 𝜕𝑥 𝑠 𝜕𝑥 𝜕𝑥
Note the dependence on the gradient of the predictive parameters. Finally,
if we preprocess the inputs by identifying an appropriate transformation
matrix P such that 𝑔(a, b) = 𝑔(Pa, Pb) = 𝑔(𝜶 , 𝜷) (8.14), then the desired
gradient of expected improvement is:
𝜕𝛼 ei 𝜕𝑔 𝜕a 𝜕𝑔 𝜕b
= P + P .
𝜕𝑥 𝜕𝜶 𝜕𝑥 𝜕𝜷 𝜕𝑥

Gradient of noisy probability of improvement

We may compute the gradient of (8.23) via several applications of the this picks up the discussion from § 8.3, p. 170
chain rule, under the (mild) assumption that the endpoints ℓ and 𝑢
correspond to unique lines (𝑎 ℓ , 𝑏 ℓ ) and (𝑎𝑢 , 𝑏𝑢 ).2 Then we have 2 If not, probability of improvement is not dif-
ferentiable as moving one of the lines at the
𝜕𝛼 pi 𝜙 (ℓ) 𝜕𝛼 pi ℓ𝜙 (ℓ) 𝜕𝛼 pi 𝜙 (−𝑢) 𝜕𝛼 pi 𝑢𝜙 (−𝑢) shared intersection will alter the probability
=− ; =− ; = ; = , of improvement only in favorable directions.
𝜕𝑎 ℓ 𝑏ℓ 𝜕𝑏 ℓ 𝑏ℓ 𝜕𝑎𝑢 𝑏𝑢 𝜕𝑏𝑢 𝑏𝑢
and we may compute the gradient with respect to the proposed observa-
tion location as
𝜕𝛼 pi 𝜕𝛼 pi 𝜕a 𝜕𝛼 pi 𝜕b
= + .
𝜕𝑥 𝜕a 𝜕𝑥 𝜕b 𝜕𝑥
The gradient with respect to a and b was computed previously (c.2).

Approximating gradient via Gauss–Hermite quadrature

We may also use numerical techniques to estimate the gradient of the this picks up the discussion from § 8.5, p. 172
acquisition function with respect to the proposed observation location:
∫
𝜕𝛼 𝜕
= Δ(𝑥, 𝑦) N (𝑦; 𝜇, 𝑠 2 ) d𝑦 (c.3)
𝜕𝑥 𝜕𝑥
Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 309
 gradients

If we assume sufficient regularity,3 then we may swap the order of ex-

3 It is sufficient for example that both Δ and
pectation and differentiation:
𝜕𝑥 be continuous in 𝑥 and 𝑦, and that the
𝜕Δ

moments of the predictive distribution of 𝑦 be ∫

𝜕𝛼 𝜕
continuous with respect to 𝑥. = Δ(𝑥, 𝑦) N (𝑦; 𝜇, 𝑠 2 ) d𝑦,
𝜕𝑥 𝜕𝑥
reducing the computation of the gradient to Gaussian expectation.
approximation with Gauss–Hermite Gauss–Hermite quadrature remains a natural choice, resulting in the
quadrature approximation
𝜕𝛼 ∑︁ 𝜕Δ
𝑛
≈ 𝑤¯ 𝑖 (𝑥, 𝑦𝑖 ), (c.4)
𝜕𝑥 𝑖=1
𝜕𝑥
accounting for dependence of {𝑦𝑖 } on which is simply the derivative of the estimator in (8.29). When using
predictive parameters this estimate we must remember that the {𝑦𝑖 } samples depend on 𝑥
through the parameters of predictive distribution (8.29). Accounting for
this dependence, the required gradient of the marginal gain at the 𝑖th
integration node 𝑧𝑖 is:
   
𝜕Δ 𝜕Δ 𝜕Δ 𝜕𝜇 𝑦𝑖 − 𝜇 𝜕𝑠
(𝑥, 𝑦𝑖 ) = + + , (c.5)
𝜕𝑥 𝜕𝑥 𝑦 𝜕𝑦 𝜕𝑥 𝑠 𝜕𝑥

where [ 𝜕Δ
𝜕𝑥 ] 𝑦 indicates the partial derivative of the updated utility when
the observed value 𝑦 is held constant.

Approximating the gradient of knowledge gradient

this picks up the discussion from § 8.6, p. 175 Special care is needed to estimate the gradient of the knowledge gradi-
ent using numerical techniques. Inspecting (c.4), we must compute the
gradient of the marginal gain
𝜕Δ 𝜕
= max 𝜇D0 (𝑥 0).
𝜕𝑥 𝜕𝑥 𝑥 0 ∈X
It is not clear how we can differentiate through the max operator in
4 p. milgrom and i. segal (2002). Envelope The- this expression. However, under mild assumptions we may appeal to the
orems for Arbitrary Choice Sets. Econometrica envelope theorem for arbitrary choice sets to proceed.4 Consider a point
70(2):583–601.
𝑥 in the interior of the domain X ⊂ ℝ𝑑, and fix an arbitrary associated
5 We assume 𝑥 is in the interior of X to avoid observation value 𝑦.5 Take any point maximizing the updated posterior
issues with defining differentiability on the
boundary. We can avoid this assumption if
mean:
the Gaussian process can be extended to all 𝑥 ∗ ∈ arg max 𝜇D0 (𝑥 0);
of ℝ𝑑. This is usually possible trivially unless 𝑥 0 ∈X
the prior mean or covariance function of the this point need not be unique. Assuming the updated posterior mean
Gaussian process is exotic. is differentiable,6 the envelope theorem states that the gradient of the
6 Sufficient conditions are that the prior mean global reward utility is equal to the gradient of the updated posterior
and covariance functions be differentiable. mean evaluated at 𝑥 ∗ :
𝜕Δ 𝜕
= 𝜇D0 (𝑥 ∗ ).
𝜕𝑥 𝜕𝑥
For the knowledge gradient, and accounting for the dependence on 𝑥 in
the locations of the 𝑦 samples (c.5), we may compute:
 
𝜕Δ 𝑦𝑖 − 𝜇 𝜕𝐾D ∗ 𝜕𝑠 ∗
(𝑥, 𝑦𝑖 ) = 𝑠 (𝑥, 𝑥𝑖 ) − 𝐾D (𝑥, 𝑥𝑖 ) ,
𝜕𝑥 𝑠3 𝜕𝑥 𝜕𝑥
310
 c.3. gradients of common acqisition functions

where 𝑥𝑖∗ maximizes the updated posterior mean corresponding to 𝑦𝑖 .

Combining this result with (c.4) yields a Gauss–Hermite approximation
to the gradient of the acquisition function. This approach requires maxi-
mizing the updated posterior mean for each sample 𝑦𝑖 . We may appeal
to standard techniques for this task, making use of the gradient (c.1)
𝜕𝜇D0
.
𝜕𝑥

Gradient of predictive entropy search acquisition function

The explicit formula (8.52) and results above allow us to compute – after this picks up the discussion from § 8.8, p. 187
investing considerable tedium – the gradient of the predictive entropy
search approximation to the mutual information (8.41). We begin by
differentiating that estimate:

𝜕𝛼 pes 1 𝜕𝑠 1 ∑︁ 1 𝜕𝑠 ∗2𝑖
𝑛
= + .
𝜕𝑥 𝑠 𝜕𝑥 2𝑛 𝑖=1 𝑠 ∗2𝑖 𝜕𝑥

Given a fixed sample 𝑥 ∗ and dropping the subscript, we may compute

(8.39):
𝜕𝑠 ∗2 𝜕𝜎∗2 𝜕𝜎𝑛2
= + .
𝜕𝑥 𝜕𝑥 𝜕𝑥
We proceed by differentiating the approximate latent predictive variance
(8.52):  
𝜕𝜎∗2 𝜕𝜍 2 2𝜍 2 − 2𝜌 𝜕𝜍 2 𝜕𝜌 (𝜍 2 − 𝜌) 2 𝜕𝛾
= − − + ,
𝜕𝑥 𝜕𝑥 𝛾 𝜕𝑥 𝜕𝑥 𝛾2 𝜕𝑥
which depends on the derivatives of the expectation propagation update
terms: "
2 #  
𝜕𝛾 1 Φ(𝑧) − 𝜙 (𝑧) 𝜕𝑚 𝜕𝜎¯ 2𝛾 𝜕𝜎¯
= 1−
2 + 𝑧 + ;
𝜕𝑥 𝛼 𝑧Φ(𝑧) + 𝜙 (𝑧) 𝜕𝑥 𝜕𝑥 𝜎¯ 𝜕𝑥
 
𝜕𝜎¯ 1 𝜕𝜍 2 𝜕𝜌
= −2 ;
𝜕𝑥 2𝜎¯ 𝜕𝑥 𝜕𝑥

and the predictive parameters (8.51):

𝜕𝑚 𝜕𝜇 𝜕k
= + (𝜶 ∗ ) > ;
𝜕𝑥 𝜕𝑥 𝜕𝑥
𝜕𝜍 2 𝜕𝜎 2 > −1 𝜕k 𝜕𝜌 𝜕𝑘 1 −1 𝜕k
= − 2k V∗ ; = − k>
∗ V∗ .
𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥

Gradient of opes acquisition function

To compute the gradient of the opes approximation, we differentiate this picks up the discussion from § 8.9, p. 192
(8.63):
𝜕𝛼 opes 1 𝜕𝑠 1 ∑︁ 𝑤𝑖 𝜕𝑠 ∗2𝑖
= + .
𝜕𝑥 𝑠 𝜕𝑥 2 𝑖 𝑠 ∗2𝑖 𝜕𝑥
Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 311
gradients

Given a fixed sample 𝑓 ∗ and dropping the subscript, we may compute

(8.63):
𝜕𝑠 ∗2 𝜕𝜎∗2 𝜕𝜎𝑛2
= + .
𝜕𝑥 𝜕𝑥 𝜕𝑥
Finally, we differentiate (8.61):
  
𝜕𝜎∗2 𝜙 (𝑧) 𝜕𝜇 𝜕𝜎 𝜙 (𝑧) 2 𝜙 (𝑧) 2
=𝜎 +𝑧 1−2 − 3𝑧 − 𝑧
𝜕𝑥 Φ(𝑧) 𝜕𝑥 𝜕𝑥 Φ(𝑧) 2 Φ(𝑧)
2
 2 
𝜕𝜎 𝜙 (𝑧) 𝜙 (𝑧)
+ 1−𝑧 − .
𝜕𝑥 Φ(𝑧) Φ(𝑧) 2

312
 ANNOTATED BIBLIOGRAPHY OF APPLICATIONS

Countless settings across science, engineering, and beyond involve free

parameters that can be tuned at will to achieve some objective. How-
ever, in many cases the evaluation of a given parameter setting can be
extremely costly, stymieing exhaustive exploration of the design space.
Of course, such a situation is a perfect use case for Bayesian opti-
mization. Through careful modeling and intelligent policy design, Bayes-
d
ian optimization algorithms can deliver impressive optimization perfor-
mance even with small observation budgets. This capability has been
demonstrated in hundreds of studies across a wide range of domains.
Here we provide a brief survey of some notable applications of Bayes-
ian optimization. The selected references are not intended to be exhaus-
tive (that would be impossible given the size of the source material!), but
rather to be representative, diverse, and good starting points for further
investigation.

chemistry and materials science

At a high level, the synthesis of everything from small molecules to
bulk materials proceeds in the same manner: initial raw materials are
combined and subjected to suitable conditions such that they transform
into a final product. Although this seems like a simple recipe, we face
massive challenges in its realization:
• We usually wish that the resulting products be useful, that is, that they
exhibit desirable properties. In drug discovery, for example, we seek
molecules exhibiting binding activity against an identified biological
target. In other settings we might seek products exhibiting favorable
optical, electronic, mechanical, thermal, and/or other properties.
• The space of possible products can be enormous and the sought after
properties exceptionally rare. For example, there are an estimated 1060
pharmacologically active molecules, only a tiny fraction of which might
exhibit binding activity against any given biological target.
• Determining the properties of a candidate molecule or material ultimately
requires synthesis and characterization in a laboratory, which can be
complicated, costly, and slow.
For these reasons, exploration of molecular or material spaces can de-
volve into a cumbersome trial-and-error search for “needles in a haystack.”
Over the past few decades, the fields of computational chemistry and computational chemistry, computational
computational materials science have developed sophisticated techniques materials science
for estimating chemical and material properties from simulation. As
accurate simulation often requires consideration of quantum mechanical
interactions, these surrogates can still be quite costly, but are nonethe-
less cheaper and more easily parallelized than laboratory experiments.
This has enabled computer-guided exploration of large molecular and
materials spaces at relatively little cost, but in many cases exhaustive

This material will be published by Cambridge University Press as Bayesian Optimization. 313
This prepublication version is free to view and download for personal use only. Not for
redistribution, resale, or use in derivative works. ©Roman Garnett 2022.
 annotated bibliography of applications

exploration remains untenable, and we must appeal to methods such as

Bayesian optimization for guidance.

Virtual screening of molecular spaces

The ability to estimate molecular properties using computational meth-
ods has enabled so-called virtual screening, where early stages of discov-
ery can be performed in silico. Bayesian optimization and active search
can dramatically increase the rate of discovery in this setting.
One challenge here is constructing predictive models for the prop-
chemoinformatics erties of molecules, which are complex structured objects. The field of
molecular fingerprints chemoinformatics has developed a number of molecular fingerprints in-
tended to serve as useful feature representations for molecules when
predicting chemical properties. Traditionally, these fingerprints were
neural fingerprints developed based on chemical intuition; however, numerous neural fin-
gerprints have emerged in recent years, designed via deep representation
1 See § 8.11, p. 199 and § 9.2, p. 209 for related learning on huge molecular databases.1
discussion.
graff, david e. et al. (2021). Accelerating high-throughput virtual screen-
ing through molecular pool-based active learning. Chemical Sci-
ence 12(22):7866–7881.
hernández-lobato, josé miguel et al. (2017). Parallel and Distributed
Thompson Sampling for Large-scale Accelerated Exploration of
Chemical Space. Proceedings of the 34th International Conference
on Machine Learning (icml 2017).
jiang, shali et al. (2017). Efficient Nonmyopic Active Search. Proceedings
of the 34th International Conference on Machine Learning (icml
2017).

De novo design
A recent trend in molecular discovery has been to learn deep genera-
tive models for molecular structures and perform optimization in the
(continuous) latent space of such a model. This enables de novo design,
where the optimization routine can propose entirely novel structures
for evaluation by feeding selected points in the latent space through the
generational procedure.1
the question of sythesizability Although appealing, a major complication with this scheme is iden-
tifying a synthetic route – if one even exists! – for proposed structures.
This issue deserves careful consideration when designing a system that
can effectively transition from virtual screening to the laboratory.

gao, wenhao et al. (2020). The Synthesizability of Molecules Proposed by

Generative Models. Journal of Chemical Information and Modeling
60(12):5714–5723.
gómez-bombarelli, rafael et al. (2018). Automatic Chemical Design
Using a Data-Driven Continuous Representation of Molecules.
acs Central Science 4(2):268–276.

314
 annotated bibliography of applications

griffiths, ryan-rhys et al. (2020). Constrained Bayesian optimization

for automatic chemical design using variational autoencoders.
Chemical Science 11(2):577–586.
korovina, ksenia et al. (2020). Chembo: Bayesian Optimization of
Small Organic Molecules with Synthesizable Recommendations.
Proceedings of the 23rd International Conference on Artificial Intel-
ligence and Statistics (aistats 2020).

Reaction optimization
Not all applications of Bayesian optimization in chemistry take the form
of optimizing chemical properties as a function of molecular structure.
For example, even once a useful molecule has been identified, there may
remain numerous parameters of its synthesis – reaction environment,
processing parameters, etc. – that can be further optimized, seeking for
example to reduce the cost and/or increase the yield of production.

shields, benjamin j. et al. (2021). Bayesian reaction optimization as a

tool for chemical synthesis. Nature 590:89–96.

Conformational search
Conformational search seeks to identify the configuration(s) of a molecule
with the lowest potential energy. Even for molecules of moderate size,
the space of possible configurations can be enormous and the potential
energy surface can exhibit numerous local minima. Interaction with
other structures such as a surface (adsorption) can further complicate adsorption
the computation of potential energy, rendering the search even more
difficult.

carr, shane f. et al. (2017). Accelerating the Search for Global Minima
on Potential Energy Surfaces using Machine Learning. Journal of
Chemical Physics 145(15):154106.
fang, lincan et al. (2021). Efficient Amino Acid Conformer Search with
Bayesian Optimization. Journal of Chemical Theory and Compu-
tation 17(3):1955–1966.
packwood, daniel (2017). Bayesian Optimization for Materials Science.
Springer–Verlag.

We may also use related Bayesian methods to map out minimal- mapping minimal-energy pathways
energy pathways between neighboring minima on a potential energy
surface in order to understand the intermediate geometry of a molecule
as it transforms from one energetically favorable state to another.

koistinen, olli-pekka et al. (2017). Nudged elastic band calculations

accelerated with Gaussian process regression. Journal of Chemical
Physics 147(15):152720.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 315

annotated bibliography of applications

Optimization of material properties and performance

Bayesian optimization has demonstrated remarkable success in acceler-
ating materials design. As in molecular design, in many cases we can
perform early screening in silico, using computational methods to approx-
imate properties of interest. The literature in this space is vast, and the
studies below are representative examples targeting a range of material
types and properties.
attia, peter m. et al. (2020). Closed-loop optimization of fast-charging
protocols for batteries with machine learning. Nature 578:397–
402.
fukazawa, taro et al. (2019). Bayesian optimization of chemical compo-
sition: A comprehensive framework and its application to 𝑅Fe12 -
type magnet compounds. Physical Review Materials 3(5):053807.
haghanifar, sajad et al. (2020). Discovering high-performance broad-
band and broad angle antireflection surfaces by machine learning.
Optica 7(7):784–789.
herbol, henry c. et al. (2018). Efficient search of compositional space for
hybrid organic–inorganic perovskites via Bayesian optimization.
npj Computational Materials 4(51).
ju, shenghong et al. (2017). Designing Nanostructures for Phonon Trans-
port via Bayesian Optimization. Physical Review x 7:021024.
miyagawa, shinsuke et al. (2021). Application of Bayesian optimiza-
tion for improved passivation performance in TiO𝑥 /SiO𝑦 /c-Si
heterostructure by hydrogen plasma treatment. Applied Physics
Express 14(2):025503.
nakamura, kensaku et al. (2021). Multi-objective Bayesian optimization
of optical glass compositions. Ceramics International 47(11):15819–
15824.
nugraha, asep sugih et al. (2020). Mesoporous trimetallic PtPdAu alloy
films toward enhanced electrocatalytic activity in methanol oxida-
tion: unexpected chemical compositions discovered by Bayesian
optimization. Journal of Materials Chemistry a 8(27):13532–13540.
osada, keiichi et al. (2020). Adaptive Bayesian optimization for epitaxial
growth of Si thin films under various constraints. Materials Today
Communications 25:1015382.
seko, atsuto et al. (2015). Prediction of Low-Thermal-Conductivity
Compounds with First-Principles Anharmonic Lattice-Dynamics
Calculations and Bayesian Optimization. Physical Review Letters
115(20):205901.

Structural search
A fundamental question in materials science is how the structure of a
material gives rise to its material properties. However, predicting the
likely structure of a material – for example by evaluating the potential
energy of plausible structures and minimizing – can be extraordinarily
difficult, as the number of possible configurations can be astronomical.

316
 annotated bibliography of applications

kiyohara, shin et al. (2016). Acceleration of stable interface structure

searching using a kriging approach. Japanese Journal of Applied
Physics 55(4):045502.
okamoto, yasuharu (2017). Applying Bayesian Approach to Combi-
natorial Problem in Chemistry. Journal of Physical Chemistry a
121(17):3299–3304.
todorović, milica et al. (2019). Bayesian inference of atomistic structure
in functional materials. npj Computational Materials 5(35).

Software for Bayesian optimization in chemistry and materials science

häse, florian et al. (2018). Phoenics: A Bayesian Optimizer for Chem-
istry. acs Central Science 4(9):1134–1145.
ueno, tsuyoshi et al. (2016). combo: An efficient Bayesian optimization
library for materials science. Materials Discovery 4:18–21.

physics
Modern physics is driven by experiments of massive scope, which have
enabled the refinement of theories of the Universe on all scales. The
complexity of these experiments – from data acquisition to the following
analysis and inference – offers many opportunities for optimization, but
the same complexity often renders optimization difficult due to large
parameter spaces and/or the expense of evaluating a particular setting.

Experimental physics
Complex physical instruments such as particle accelerators offer the
capability of extraordinarily fine tuning through careful setting of their
control parameters. In some cases, these parameters can be altered on-the-
fly during operation, resulting in a huge space of possible configurations.
Bayesian optimization can help accelerate the tuning process.

duris, j. et al. (2020). Bayesian Optimization of a Free-Electron Laser.

Physical Review Letters 124(12):124801.
roussel, ryan et al. (2021). Multiobjective Bayesian optimization for
online accelerator tuning. Physical Review Accelerators and Beams
24(6):062801.
shalloo, r. j. et al. (2020). Automation and control of laser wakefield
accelerators using Bayesian optimization. Nature Communications
11:6355.
wigley, p. b. et al. (2016). Fast machine-learning online optimization of
ultra-cold-atom experiments. Scientific Reports 6:25890.

Inverse problems in physics

The goal of a inverse problem is to determine the free parameters of a 2 This model may not be probabilistic but rather
generative model2 from observations of its output. Inverse problems a complex simulation procedure.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 317

annotated bibliography of applications

are pervasive in physics – many physical models feature parameters

(physical constants) that cannot be determined from first principles.
However, we can infer these parameters experimentally by comparing
the predictions of the model with the behavior of relevant observations.
In this context, determining the predictions of the model given a
forward map setting of its parameters is called the forward map. The inverse problem
is then the task of “inverting” this map: searching the parameter space
for the settings in the greatest accord with observed data.
Bayesian analog We can draw parallels here with Bayesian inference, where the for-
ward map is characterized by the likelihood, which determines the dis-
tribution of observed data given the parameters. The Bayesian answer
to the inverse problem is then encapsulated by the posterior distribution
of the parameters given the data, which identifies the most plausible
parameters in light of the observations.
For complex physical models, even the forward map can be excep-
tionally expensive to compute, making it difficult to completely explore
its parameter space. For example, the forward map of a cosmological
model may require simulating the evolution of an entire universe at a
fine enough resolution to observe its large scale structure. Exhaustively
traversing the space of cosmological parameters and comparing with
observed structure would be infeasible, but progress may be possible
with careful guidance. Bayesian optimization has proven useful to this
end on a range of difficult inverse problems.

ilten, p. et al. (2017). Event generator tuning using Bayesian optimization.

Journal of Instrumentation 12(4):P04028.
leclercq, florent (2018). Bayesian optimization for likelihood-free
cosmological inference. Physical Review d 98(6):063511.
rogers, keir k. et al. (2019). Bayesian emulator optimisation for cosmol-
ogy: application to the Lyman-alpha forest. Journal of Cosmology
and Astroparticle Physics 2019(2):031.
vargas-hernández, r. a. et al. (2019). Bayesian optimization for the
inverse scattering problem in quantum reaction dynamics. New
Journal of Physics 21(2):022001.

biological sciences and engineering

Biological systems are extraordinarily complex. Obtaining experimental
measurements to shed light on the behavior of these systems can be
difficult, slow, and expensive, and the resulting data can be corrupted
with significant noise. Efficient experimental design is thus critical to
make progress, and Bayesian optimization is a natural tool to consider.

li, yan et al. (2018). A Knowledge Gradient Policy for Sequencing Exper-
iments to Identify the Structure of rna Molecules Using a Sparse
Additive Belief Model. informs Journal on Computing 30(4):625–
786.

318
 annotated bibliography of applications

lorenz, romy et al. (2018). Dissociating frontoparietal brain networks

with neuroadaptive Bayesian optimization. Nature Communica-
tions 9:1227.
nikitin, artyom et al. (2019). Bayesian optimization for seed germina-
tion. Plant Methods 15:43.

Inverse problems in the biological sciences

Challenging inverse problems (see above) are pervasive in biology. Effec-
tive modeling of biological systems often requires complicated, nonlinear
models with numerous free parameters. Bayesian optimization can help
guide the search for the parameters offering the best explanation of
observed data.

dokoohaki, hamze et al. (2018). Use of inverse modelling and Bayes-

ian optimization for investigating the t effect of biochar on soil
hydrological properties. Agricultural Water Management 2018:
268–274.
thomas, marcus et al. (2018). A method for efficient Bayesian optimiza-
tion of self-assembly systems from scattering data. bmc Systems
Biology 12:65.
ulmasov, doniyor et al. (2016). Bayesian Optimization with Dimension
Scheduling: Application to Biological Systems. Computer Aided
Chemical Engineering 38:1051–1056.

Gene and protein design

The tools of modern biology enable the custom design of genetic se-
quences and even entire proteins, which we can – at least theoretically
– tailor as we see fit. It is natural to pose both gene and protein design
in terms of optimizing some figure of merit over a space of alternatives.
However, in either case we face an immediate combinatorial explosion
in the number of possible genetic or amino acid sequences we might
consider. Bayesian optimization has shown promise for overcoming this
obstacle through careful experimental design. A running theme through-
out recent efforts is the training of generative models for gene/protein
sequences, after which we may optimize in the continuous latent space 3 See § 8.11, p. 199 and § 9.2, p. 209 for related
in order to sidestep the need for combinatorial optimization.3 discussion.

gonzález, javier et al. (2015). Bayesian Optimization for Synthetic Gene

Design. Bayesian Optimization: Scalability and Flexibility Work-
shop (BayesOpt 2015), Conference on Neural Information Processing
Systems (neurips 2015).
hie, brian l. et al. (2021). Adaptive machine learning for protein engi-
neering. arXiv: 2106.05466 [q-bio.QM].
romero, philip a. et al. (2013). Navigating the protein fitness landscape
with Gaussian processes. Proceedings of the National Academy of
Sciences 110(3):e193–e201.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 319

annotated bibliography of applications

sinai, sam et al. (2020). A primer on model-guided exploration of fitness

landscapes for biological sequence design. arXiv: 2010 . 10614
[q-bio.QM].
yang, kevin k. et al. (2019). Machine-learning-guided directed evolution
for protein engineering. Nature Methods 16:687–694.

plant breeding A related problem is faced in modern plant breeding, where the
goal is to develop plant varieties with desirable characteristics: yield,
drought or pest resistance, etc. Plant phenotyping is an inherently slow
process, as we must wait for planted seeds to germinate and grow until
sufficiently mature that traits can be measured. This slow turnover rate
makes it impossible to fully explore the space of possible genotypes, the
genetic information that (in combination with other factors including
environment and management) gives rise to the phenotypes we seek
to optimize. Modern plant breeding uses genetic sequencing to guide
the breeding process by building models to predict phenotype from
genotype and using these models in combination with large gene banks
to inform the breeding process. A challenge in this approach is that the
space of possible genotypes is huge, but Bayesian optimization can help
accelerate the search.

tanaka, ryokei et al. (2018). Bayesian optimization for genomic selection:

a method for discovering the best genotype among a large number
of candidates. Theoretical and Applied Genetics 131(1):93–105.

Biomedical engineering
Difficult optimization problems are pervasive in biomedical engineering
due to the complexity of the systems involved and the often considerable
cost of gathering data, whether through studies with human subjects,
complicated laboratory testing, and/or nontrivial simulation.

colopy, glen wright et al. (2018). Bayesian Optimization of Person-

alized Models for Patient Vital-Sign Monitoring. ieee Journal of
Biomedical and Health Informatics 22(2):301–310.
ghassemi, mohammad et al. (2014). Global Optimization Approaches for
Parameter Tuning in Biomedical Signal Processing: A Focus of
Multi-scale Entropy. Computing in Cardiology 41(12–2):993–996.
kim, gilhwan et al. (2021). Using Bayesian Optimization to Identify
Optimal Exoskeleton Parameters Targeting Propulsion Mechanics:
A Simulation Study. bioRxiv: 2021.01.14.426703.
kim, myunghee et al. (2017). Human-in-the-loop Bayesian optimization
of wearable device parameters. plos one 12(9):e0184054.
olofsson, simon et al. (2019). Bayesian Multiobjective Optimisation
With Mixed Analytical and Black-Box Functions: Application to
Tissue Engineering. ieee Transactions on Biomedical Engineering
66(3):727–739.

320
 annotated bibliography of applications

robotics
Robotics is fraught with difficult optimization problems. A robotic plat-
form may have numerous tunable parameters influencing its behavior,
and the dependence of its performance on these parameters may be
highly complex. Further, empirical evaluation can be difficult – real-world
experiments must proceed in real time, and there may be a considerable
set-up time between experiments.

bansal, somil et al. (2017). Goal-Driven Dynamics Learning via Bayesian

Optimization. Proceedings of the 56th Annual ieee Conference on
Decision and Control (cdc 2017.
calandra, roberto et al. (2016). Bayesian optimization for learning gaits
under uncertainty: An experimental comparison on a dynamic
bipedal walker. Annals of Mathematics and Artificial Intelligence
76(1–2):5–23.
junge, kai et al. (2020). Improving Robotic Cooking Using Batch Bayes-
ian Optimization. ieee Robotics and Automation Letters 5(2):760–
765.
martinez-cantin, ruben et al. (2009). A Bayesian exploration–exploita-
tion approach for optimal online sensing and planning with a
visually guided mobile robot. Autonomous Robotics 27(2):93–103.

In some cases we may be able to accelerate optimization by augment- multifidelity optimization: § 11.5, p. 263
ing real-world evaluation with simulation in a multifidelity setup.

marco, alonso et al. (2017). Virtual vs. Real: Trading Off Simulations and
Physical Experiments in Reinforcement Learning with Bayesian
Optimization. Proceedings of the 2017 ieee International Conference
on Robotics and Automation (icra 2017).
rai, akshara et al. (n.d.). Using Simulation to Improve Sample-Efficiency
of Bayesian Optimization for Bipedal Robots. Journal of Machine
Learning Research 20(49).

Modeling for robotics

Modeling robot performance as a function of its parameters can be diffi-
cult due to nominally high-dimensional parameter spaces and the poten-
tial for context-dependent nonstationarity. This difficulty has motivated
sophisticated modeling approaches for addressing these issues.

jaqier, noémie et al. (2019). Bayesian Optimization Meets Riemannian

Manifolds in Robot Learning. Proceedings of the 3rd Conference
on Robot Learning (corl 2019).
martinez-cantin, ruben (2017). Bayesian Optimization with Adaptive
Kernels for Robot Control. Proceedings of the 2017 ieee Interna-
tional Conference on Robotics and Automation (icra 2017).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 321

annotated bibliography of applications

yuan, kai et al. (2019). Bayesian Optimization for Whole-Body Con-

trol of High-Degree-of-Freedom Robots Through Reduction of
Dimensionality. ieee Robotics and Automation Letters 4(3):2268–
2275.

Safe and robust optimization

robust optimization: § 11.9, p. 277 A complication faced in some robotic optimization settings is in ensuring
that the evaluated parameters are both robust (that is, that performance
is not overly sensitive to minor perturbations in the parameters) and safe
(that is, that there is no chance of catastrophic failure in the robotic plat-
form). We may address these concerns in the design of the optimization
policy. For example, we might realize robustness by redefining the utility
utility functions: chapter 6, p. 109 function in terms of the expected performance of perturbed parameters,
constrained optimization: § 11.2, p. 249 and we might realize safety by incorporating appropriate constraints.

berkenkamp, felix et al. (2021). Bayesian optimization with safety con-

straints: safe and automatic parameter tuning in robotics. Machine
Learning Special Issue on Robust Machine Learning.
garcía-barcos, javier et al. (2021). Robust policy search for robot navi-
gation. ieee Robotics and Automation Letters 6(2):2389–2396.
nogueira, josé et al. (2016). Unscented Bayesian Optimization for Safe
Robot Grasping. Proceedings of the 2016 ieee/rsj International
Conference on Intelligent Robots and Systems (iros 2016).

Adversarial attacks
A specific safety issue to consider in robotic platforms incorporating
deep neural networks is the possibility of adversarial attacks that may be
able to alter the environment so as to induce unsafe behavior. Bayesian
optimization has been applied to the efficient construction of adversarial
attacks; this capability can in turn be used during the design phase
seeking to build robotic controllers that are robust to such attack.

boloor, adith et al. (2020). Attacking vision-based perception in end-to-

end autonomous driving models. Journal of Systems Architecture
110:101766.
ghosh, shromona et al. (2018). Verifying Controllers Against Adversarial
Examples with Bayesian Optimization. Proceedings of the 2018
ieee International Conference on Robotics and Automation (icra
2018).

reinforcement learning
The main challenge in reinforcement learning is that agents can only
evaluate decision policies through trial-and-error: by following a pro-
posed policy and observing the resulting reward. When the system is
complex, evaluating a given policy may be costly in terms of time or

322
 annotated bibliography of applications

other resources, and when the space of potential policies is large, careful
exploration is necessary.
Policy search is one approach to reinforcement learning that models
the problem of policy design in terms of optimization. We enumerate
a space of potential policies (for example, via parameterization) and
seek to maximize the expected cumulative reward over the policy space.
This abstraction allows the straightforward use of Bayesian optimiza-
tion to construct a series of policies seeking to balance exploration and
exploitation of the policy space. A recurring theme in this research is
the careful use of simulators, when available, to help guide the search in
a multifidelity setup. multifidelity optimization: § 11.5, p. 263
Reinforcement learning is a central concept in robotics, and many
of the papers cited in the previous section represent applications of
policy search to particular robotic systems. The citations below concern
reinforcement learning in a broader context.

feng, qing et al. (2020). High-Dimensional Contextual Policy Search

with Unknown Context Rewards using Bayesian Optimization.
Advances in Neural Information Processing Systems 33 (neurips
2020).
letham, benjamin et al. (2019). Bayesian Optimization for Policy Search
via Online–Offline Experimentation. Journal of Machine Learning
Research 20(145):1–30.
wilson, aaron et al. (2014). Using Trajectory Data to Improve Bayesian
Optimization for Reinforcement Learning. Journal of Machine
Learning Research 15(8):253–282.

Bayesian optimization has also proven useful for variations on the

classical reinforcement learning problem, such as the multi-agent set-
ting or the so-called inverse reinforcement learning setting, where an
agent must reconstruct an unknown reward function from observing
demonstrations from other agents.

balakrishnan, sreejith et al. (2020). Efficient Exploration of Reward

Functions in Inverse Reinforcement Learning via Bayesian Opti-
mizaiton. Advances in Neural Information Processing Systems 33
(neurips 2020).
dai, zhongxiang et al. (2020). r2-b2: Recursive Reasoning-Based Bayes-
ian Optimization for No-Regret Learning in Games. Proceedings
of the 37th International Conference on Machine Learning (icml
2020).
imani, mahdi et al. (2021). Scalable Inverse Reinforcement Learning
Through Multifidelity Bayesian Optimization. ieee Transactions
on Neural Networks and Learning Systems.

civil engineering
The optimization of large-scale critical systems such as power, trans-
portation, water distribution, and sensor networks can be difficult due

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 323

annotated bibliography of applications

to complex dynamics and the considerable expense of reconfiguration.

Optimization can sometimes be aided via computer simulation, but the
configuration spaces involved can nonetheless be huge, precluding ex-
haustive search.

baheri, ali et al. (2017). Altitude Optimization of Airborne Wind Energy

Systems: A Bayesian Optimization Approach. Proceedings of the
2017 American Control Conference (acc 2017).
cornejo-bueno, l. et al. (2018). Bayesian optimization of a hybrid system
for robust ocean wave features prediction. Neurocomputing 275:
818–828.
garnett, r. et al. (2010). Bayesian Optimization for Sensor Set Selec-
tion. Proceedings of the 9th acm/ieee International Conference on
Information Processing in Sensor Networks (ipsn 2010).
gramacy, robert b. et al. (2016). Modeling an Augmented Lagrangian
for Blackbox Constrained Optimization. Technometrics 58(1):1–11.
hickish, bob et al. (2020). Investigating Bayesian Optimization for rail
network optimization. International Journal of Rail Transporation
8(4):307–323.
kopsiaftis, george et al. (2019). Gaussian Process Regression Tuned by
Bayesian Optimization for Seawater Intrusion Prediction. Com-
putational Intelligence and Neuroscience 2019:2859429.
marchant, roman et al. (2012). Bayesian Optimisation for Intelligent
Environmental Monitoring. Proceedings of the 2012 ieee/rsj Inter-
national Conference on Intelligent Robots and Systems (iros 2012).

Structural engineering
The following study applied Bayesian optimization in a structural en-
gineering setting: tuning the hyperparameters of a real-time predictive
model for the typhoon-induced response of a ∼1000 m real-world bridge.

zhang, yi-ming et al. (2021). Probabilistic Framework with Bayesian Op-

timization for Predicting Typhoon-Induced Dynamic Responses
of a Long-Span Bridge. Journal of Structural Engineering 147(1):
04020297.

electrical engineering
Over the past few decades, sophisticated tools have enabled the automa-
tion of many aspects of digital circuit design, even for extremely large
circuits. However, the design of analog circuits remains largely a man-
ual process. As circuits grow increasingly complex, the optimization of
analog circuits (for example, to minimize power consumption subject to
performance constraints) is becoming increasingly more difficult. Even
computer simulation of complex analog circuits can entail significant
cost, so careful experimental design is imperative for exploring the design
space. Bayesian optimization has proven effective in this regard.

324
 annotated bibliography of applications

chen, peng (2015). Bayesian Optimization for Broadband High-Efficiency

Power Amplifier Designs. ieee Transactions on Microwave Theory
and Techniques 63(12):4263–4272.
fang, yaoran et al. (2018). A Bayesian Optimization and Partial Element
Equivalent Circuit Approach to Coil Design in Inductive Power
Transfer Systems. Proceedings of the 2018 ieee pels Workshop on
Emerging Technologies: Wireless Power Transfer (wow 2018).
liu, mingjie et al. (2020). Closing the Design Loop: Bayesian Optimiza-
tion Assisted Hierarchical Analog Layout Synthesis. Proceedings
of the 57th acm/ieee Design Automation Conference (dac 2020).
lyu, wenlong et al. (2018). An Efficient Bayesian Optimization Approach
for Automated Optimization of Analog Circuits. ieee Transactions
on Circuits and Systems–i: Regular Papers 65(6):1954–1967.
torun, hakki mert et al. (2018). A Global Bayesian Optimization Al-
gorithm and Its Application to Integrated System Design. ieee
Transactions on Very Large Scale Integration (vlsi) Systems 26(4):
792–802.

mechanical engineering
The following study applied Bayesian optimization in a mechanical
engineering setting: tuning the parameters of a welding process (via
slow and expensive real-world experiments) to maximize weld quality.

sterling, dillon et al. (2015). Welding Parameter Optimization Based on

Gaussian Process Regression Bayesian Optimization Algorithm.
Proceedings of the 2015 ieee International Conference on Automation
Science and Engineering (case 2015).

Aerospace engineering
Nuances in airfoil design can have significant impact on aerodynamic
performance, improvements in which can in turn lead to considerable
cost savings from increased fuel efficiency. However, airfoil optimization
is challenging due to the large design spaces involved and the nontrivial
cost of evaluating a proposed configuration. Empirical measurement
requires constructing an airfoil and testing its performance in a wind
chamber. This process is too slow to explore the configuration space
effectively, but we can simulate the process via computational fluid 4 There are many parallels with this situa-
dynamics. These computational surrogates are still fairly costly due to tion and that faced in (computational) chem-
istry and materials science, where quantum-
the need to numerically solve nonlinear partial differential equations mechanical simulation also requires numer-
(the Navier–Stokes equations) at a sufficiently fine resolution, but they ically solving a partial differential equation
are nonetheless cheaper and more easily parallelized than experiments.4 (the Schrödinger equation).
Bayesian optimization can accelerate airfoil optimization via careful
and cost-aware experimental design. One important idea here that can
lead to significant computational savings is multifidelity modeling and multifidelity optimization: § 11.5, p. 263
optimization. It is relatively easy to control the cost–fidelity tradeoff
in computational fluid dynamics simulations by altering its resolution

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 325

annotated bibliography of applications

accordingly. This allows to rapidly explore the design space with cheap-
but-rough simulations, then refine the most promising regions.

chaitanya, paruchuri et al. (2020). Bayesian optimisation for low-

noise aerofoil design with aerodynamic constraints. International
Journal of Aeroacoustics 20(1–2):109–129.
forrester alexander, i. j. et al. (2009). Recent advances in surrogate-
based optimization. Progress in Aerospace Sciences 45(1–3):50–79.
hebbal, ali et al. (2019). Multi-objective optimization using Deep Gaus-
sian Processes: Application to Aerospace Vehicle Design.
lam, remi r. et al. (2018). Advances in Bayesian Optimization with Ap-
plications in Aerospace Engineering.
priem, rémy et al. (2020). An efficient application of Bayesian optimiza-
tion to an industrial mdo framework for aircraft design. Proceed-
ings of the 2020 aiaa Aviation Forum.
reisenthel, patrick h. et al. (2011). A Numerical Experiment on Al-
locating Resources Between Design of Experiment Samples and
Surrogate-Based Optimization Infills. Proceedings of the 2011
aiaa/asme/asce/ahs/ asc Structures, Structural Dynamics and Ma-
terials Conference.
zheng, hongyu et al. (2020). Multifidelity kinematic parameter opti-
mization of a flapping airfoil. Physical Review e 101(1):013107.

Automobile engineering
Bayesian optimization has also proven useful in automobile engineering.
Automobile components and subsystems can have numerous tunable pa-
rameters affecting performance, and evaluating a given configuration can
be complicated due to complex interactions among vehicle components.
Bayesian optimization has shown success in this setting.

liessner, roman et al. (2019). Simultaneous Electric Powertrain Hard-

ware and Energy Management Optimization of a Hybrid Electric
Vehicle Using Deep Reinforcement Learning and Bayesian Opti-
mization. Proceedings of the 2019 ieee Vehicle Power and Propulsion
Conference (vppc 2019).
neumann-brosig, matthias et al. (2020). Data-Efficient Autotuning
With Bayesian Optimization: An Industrial Control Study. ieee
Transactions on Control Systems Technology 28(3):730–740.
thomas, sinnu susan et al. (2019). Designing MacPherson Suspension
Architectures using Bayesian Optimization. Proceedings of the
28th Belgian Dutch Conference on Machine Learning (Benelearn
2019).

algorithm configuration, hyperparameter tuning, and automl

Complex algorithms and software pipelines can have numerous param-
eters influencing their performance, and determining the optimal con-

326
 annotated bibliography of applications

figuration for a given situation can require a significant trial-and-error.

Bayesian optimization has demonstrated remarkable success on a range
of problems under the umbrella of algorithm configuration.
In the most general setting, we may simply model algorithm perfor-
mance as a black box. A challenge here is dealing with high-dimensional
parameter spaces that may have complex structure, such as conditional
parameters that only become relevant depending on the settings of other
parameters. This can pose a challenge for Gaussian process models, but alternatives to gps, including random forests:
alternatives such as random forests can perform admirably. § 8.11, p. 196

dalibard, valentin et al. (2017). boat: Building Auto-Tuners with Struc-

tured Bayesian Optimization. Proceedings of the 26th International
Conference on World Wide Web (www 2017).
gonzalvez, joan et al. (2019). Financial Applications of Gaussian Pro-
cesses and Bayesian Optimization. arXiv: 1903.04841 [q-fin.PM].
hoos, holger h. (2012). Programming by Optimization. Communications
of the acm 55(2):70–80.
hutter, frank et al. (2011). Sequential Model-Based Optimization for
General Algorithm Configuration. Proceedings of the 5th Learning
and Intelligent Optimization Conference (lion 5).
kunjir, mayuresh (2019). Guided Bayesian Optimization to AutoTune
Memory-based Analytics. Proceedings of the 35th ieee Interna-
tional Conference on Data Engineering Workshops (icdew 2019).
słowik, agnieszka et al. (2019). Bayesian Optimisation for Heuristic
Configuration in Automated Theorem Proving. Proceedings of
the 5th and 6th Vampire Workshops (Vampire 2019).
vargas-hernández, r. a. (2020). Bayesian Optimization for Calibrat-
ing and Selecting Hybrid-Density Functional Models. Journal of
Physical Chemistry a 124(20):4053–4061.

Hyperparameter tuning of machine learning algorithms

The task of configuring machine learning algorithms in particular is
known as hyperparameter tuning. Hyperparameter tuning is especially
challenging in the era of deep learning, due to the often considerable cost
of training and validating a proposed configuration. However, Bayesian
optimization has proven effective at this task, and the results of a recent
black-box optimization competition (run by turner et al. below) focus-
ing on optimization problems from machine learning soundly established
its superiority to alternatives such as random search.

qitadamo, andrew et al. (2017). Bayesian Hyperparameter Optimiza-

tion for Machine Learning Based eqtl Analysis. Proceedings of the
8th acm International Conference on Bioinformatics, Computational
Biology, and Health Informatics (bcb 2017).
snoek, jasper et al. (2012). Practical Bayesian Optimization of Machine
Learning Algorithms. Advances in Neural Information Processing
Systems 25 (neurips 2012).

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 327

annotated bibliography of applications

turner, ryan et al. (2021). Bayesian Optimization is Superior to Random

Search for Machine Learning Hyperparameter Tuning: Analysis
of the Black-Box Optimization Challenge 2020. Proceedings of
the Neurips 2020 Competition and Demonstration Track.
yogatama, dani et al. (2015). Bayesian Optimization of Text Representa-
tions. Proceedings of the 2015 Conference on Empirical Methods in
Natural Language Processing (emnlp 2015).
tuning samplers, etc. In addition to tuning machine learning algorithms to maximize pre-
dictive performance, Bayesian optimization can also tune the parameters
of other algorithms for learning and inference such as Monte Carlo
samplers.
hamze, firas et al. (2013). Self-Avoiding Random Dynamics on Integer
Complex Systems. acm Transactions on Modeling and Computer
Simulation 23(1):9.
mahendran, nimalan et al. (2010). Adaptive mcmc with Bayesian Op-
timization. Proceedings of the 13th International Conference on
Artificial Intelligence and Statistics (aistats 2010).
One recent high-profile use of Bayesian optimization was in tuning
the hyperparameters of DeepMind’s AlphaGo agent. Bayesian optimiza-
tion was able to improve AlphaGo’s self-play win rate from one-half
of games to nearly two-thirds, and the version tuned with Bayesian
optimization was used in the final match against Lee Sedol.
chen, yutian et al. (2018). Bayesian Optimization in AlphaGo. arXiv:
1812.06855 [cs.LG].

Automated machine learning

The goal of automated machine learning (automl) is to develop automated
procedures for machine learning tasks in order to boost efficiency and
open the power of machine learning to a wider audience. Hyperparam-
eter tuning is one particular instance of this overall vision, but we can
also consider the automation of other aspects of machine learning.
model assessment, selection, and averaging: In model selection, for example, we seek to tune not only the hyper-
chapter 4, p. 67 parameters of a machine learning model but also the structure of the
model itself. One notable special case is neural architecture search, the
neural architecture search optimization of neural network architecture. Such problems are particu-
larly difficult due to the discrete, structured nature of the search spaces
involved. However, with careful modeling and acquisition strategies,
Bayesian optimization becomes a feasible solution.

jin, haifeng et al. (2019). Auto-Keras: An Efficient Neural Architecture

Search System. Proceedings of the 25th acm sigkdd International
Conference on Knowledge Discovery and Data Mining (kdd 2019).
kandasamy, kirthevasan et al. (2018). Neural Architecture Search with
Bayesian Optimisation and Optimal Transport. Advances in Neu-
ral Information Processing Systems 31 (neurips 2018).

328
 annotated bibliography of applications

malkomes, gustavo et al. (2016). Bayesian optimization for automated

model selection. Advances in Neural Information Processing Sys-
tems 29 (neurips 2016).
white, colin et al. (2021). bananas: Bayesian Optimization with Neural
Architectures for Neural Architecture Search. Proceedings of 35th
aaai Conference on Artificial Intelligence (aaai 2021).
Researchers have even considered using Bayesian optimization for
dynamic model selection during Bayesian optimization itself to realize
fully autonomous and robust optimization routines.
malkomes, gustavo et al. (2018). Automating Bayesian optimization with
Bayesian optimization. Advances in Neural Information Processing
Systems 31 (neurips 2018).

Optimization of entire machine learning pipelines

We may also consider the joint optimization of entire machine learn-
ing pipelines, starting from raw data and automatically constructing
a bespoke machine learning system. The space of possible pipelines is
absolutely enormous, as we must consider preprocessing steps such as
feature selection and engineering in addition to model selection and
hyperparameter tuning. Nonetheless, Bayesian optimization has proven
up to the task.
Building a working automl system of this scope entails a number of
subtle design questions: the space of pipelines to consider, the allocation
of training resources to proposed pipelines (which may vary enormously
in their complexity), and the modeling of performance as a function of
dataset and pipeline, to name a few. All of these questions have received
careful consideration in the literature, and the following reference is an
excellent entry point to that body of work.
feurer, matthias et al. (2020). Auto-Sklearn 2.0: The Next Generation.
arXiv: 2007.04074 [cs.LG].

adaptive human–computer interfaces

Adaptive human–computer interfaces seek to tailor themselves on-the-
fly to suit the preferences of the user and/or the system provider. For
example:
• a content provider may wish to learn user preferences to ensure recom-
mended content is relevant,
• a website depending on ad revenue may wish to tune their interface and
advertising placement algorithms to maximize ad revenue,
• a computer game may seek to adjust its difficulty dynamically to keep
the player engaged, or
• a data-visualization system may seek to infer the user’s goals in interact-
ing with a dataset and customize the presentation of data accordingly.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 329

annotated bibliography of applications

The design of such a system can be challenging, as the space of possi-

ble user preferences and/or the space of possible algorithmic settings
can be large, and the optimal configuration may change over time or
depend on other context. Further, we may only assess the utility of a
given interface configuration through user interaction, which is a slow,
cumbersome, and noisy channel from which to glean information. Fi-
nally, we face the additional challenge that if we are not careful, the
user may become annoyed and simply abandon the platform altogether!
Nonetheless, Bayesian optimization has shown success in tuning adap-
tive interfaces and in related problems such as preference optimization,
a/b testing, etc.

brochu, eric et al. (2010). A Bayesian Interactive Optimization Approach

to Procedural Animation Design. Proceedings of the 2010 acm
siggraph/Eurographics Symposium on Computer Animiation (sca
2010).
brochu, eric et al. (2015). Active Preference Learning with Discrete
Choice Data. Advances in Neural Information Processing Systems
20 (neurips 2007).
gonzález, javier et al. (2017). Preferential Bayesian Optimization. Pro-
ceedings of the 34th International Conference on Machine Learning
(icml 2017).
kadner, florian et al. (2021). AdaptiFont: Increasing Individuals’ Read-
ing Speed with a Generative Font Model and Bayesian Optimiza-
tion. Proceedings of the 2021 chi Conference on Human Factors in
Computing Systems (chi 2021).
khajah, mohammad m. et al. (2016). Designing Engaging Games Using
Bayesian Optimization. Proceedings of the 2016 chi Conference on
Human Factors in Computing Systems (chi 2016).
letham, benjamin et al. (2019). Constrained Bayesian Optimization with
Noisy Experiments. Bayesian Analysis 14(2):495–519.
monadjemi, shayan et al. (2020). Active Visual Analytics: Assisted Data
Discovery in Interactive Visualizations via Active Search. arXiv:
2010.08155 [cs.HC].

330
BIBLIOGRAPHY

abbasi-yadkori, yasin (2012). Online Learning for Linearly Parameterized Control Prob-
lems. PhD thesis. University of Alberta.
acerbi, luigi (2018). Variational Bayesian Monte Carlo. Advances in Neural Information
Processing Systems 31 (neurips 2018), pp. 8213–8223.
acerbi, luigi and wei ji ma (2017). Practical Bayesian Optimization for Model Fitting
with Bayesian Adaptive Direct Search. Advances in Neural Information Processing
Systems 30 (neurips 2017), pp. 1836–1846.
adams, ryan prescott, iain murray, and david j. c. mackay (2009). Tractable Nonpara-
metric Bayesian Inference in Poisson Processes with Gaussian Process Intensities.
Proceedings of the 26th International Conference on Machine Learning (icml 2009),
pp. 9–16.
adler, robert j. and jonathan e. taylor (2007). Random Fields and Geometry. Springer
Monographs in Mathematics. Springer–Verlag.
agrawal, rajeev (1995). The Continuum-Armed Bandit Problem. siam Journal on Control
and Optimization 33(6):1926–1951.
agrawal, shipra and navin goyal (2012). Analysis of Thompson Sampling for the Multi-
armed Bandit Problem. Proceedings of the 25th Annual Conference on Learning
Theory (colt 2012). Vol. 23. Proceedings of Machine Learning Research, pp. 39.1–
39.26.
álvarez, mauricio a., lorenzo rosasco, and neil d. lawrence (2012). Kernels for Vector-
Valued Functions: A Review. Foundations and Trends in Machine Learning 4(3):195–
266.
arcones, miguel a. (1992). On the arg max of a Gaussian process. Statistics & Probability
Letters 15(5):373–374.
auer, peter, nicolò cesa-bianchi, and paul fischer (2002). Finite-time Analysis of the
Multiarmed Bandit Problem. Machine Learning 47(2–3):235–256.
balandat, maximilian, brian karrer, daniel r. jiang, samuel daulton, benjamin
letham, andrew gordon wilson, et al. (2020). BoTorch: A Framework for Effi-
cient Monte-Carlo Bayesian Optimization. Advances in Neural Information Process-
ing Systems 33 (neurips 2020), pp. 21524–21538.
baptista, ricardo and matthias poloczek (2018). Bayesian Optimization of Combinato-
rial Structures. Proceedings of the 35th International Conference on Machine Learning
(icml 2018). Vol. 80. Proceedings of Machine Learning Research, pp. 462–471.
bather, john (1996). A Conversation with Herman Chernoff. Statistical Science 11(4):335–
350.
belakaria, syrine, aryan deshwal, and janardhan rao doppa (2019). Max-value En-
tropy Search for Multi-Objective Bayesian Optimization. Advances in Neural Infor-
mation Processing Systems 32 (neurips 2019), pp. 7825–7835.
bellman, richard (1952). On the Theory of Dynamic Programming. Proceedings of the
National Academy of Sciences 38(8):716–719.
bellman, richard (1957). Dynamic Programming. Princeton University Press.
berger, james o. (1985). Statistical Decision Theory and Bayesian Analysis. 2nd ed. Springer
Series in Statistics. Springer–Verlag.
bergstra, james, rémi bardenet, yoshua bengio, and balázs kégl (2011). Algorithms for
Hyper-Parameter Optimization. Advances in Neural Information Processing Systems
24 (neurips 2011), pp. 2546–2554.

This material will be published by Cambridge University Press as Bayesian Optimization. This prepublication 331
version is free to view and download for personal use only. Not for redistribution, resale, or use in derivative
works. ©Roman Garnett 2022.
bibliography

bergstra, james and yoshua bengio (2012). Random Search for Hyper-Parameter Opti-
mization. Journal of Machine Learning Research 13:281–305.
berkenkamp, felix, angela p. schoelling, and andreas krause (2019). No-Regret Bayes-
ian Optimization with Unknown Hyperparameters. Journal of Machine Learning
Research 20(50):1–24.
berry, donald a. and bert fristedt (1985). Bandit Problems: Sequential Allocation of
Experiments. Monographs on Statistics and Applied Probability. Chapman & Hall.
bertsekas, dimitri p. (2017). Dynamic Programming and Optimal Control. 4th ed. Vol. 1.
Athena Scientific.
bochner, s (1933). Monotone Funktionen, Stieltjessche Integrale und harmonische Analyse.
Mathematische Annalen 108:378–410.
bogunovic, ilija, andreas krause, and jonathan scarlett (2020). Corruption-Tolerant
Gaussian Process Bandit Optimization. Proceedings of the 23rd International Confer-
ence on Artificial Intelligence and Statistics (aistats 2020). Vol. 108. Proceedings of
Machine Learning Research, pp. 1071–1081.
bogunovic, ilija, jonathan scarlett, stefanie jegelka, and volkan cevher (2018).
Adversarially Robust Optimization with Gaussian Processes. Advances in Neural
Information Processing Systems 31 (neurips 2018), pp. 5760–5770.
box, g. e. p. (1954). The Exploration and Exploitation of Response Surfaces: Some General
Considerations and Examples. Biometrics 10(1):16–60.
box, george e. p., j. stuart hunter, and william g. hunter (2005). Statistics for Experi-
menters: Design, Innovation, and Discovery. 2nd ed. Wiley Series in Probability and
Statistics. John Wiley & Sons.
box, g. e. p. and k. b. wilson (1951). On the Experimental Attainment of Optimum Condi-
tions. Journal of the Royal Statistical Society Series B (Methodological) 13(1):1–45.
box, g. e. p. and p. v. youle (1954). The Exploration and Exploitation of Response Surfaces:
An Example of the Link Between the Fitted Surface and the Basic Mechanism of
the System. Biometrics 11(3):287–323.
breiman, leo (2001). Random Forests. Machine Learning 45(1):5–32.
brent, richard p. (1973). Algorithms for Minimization Without Derivatives. Prentice–Hall
Series in Automatic Computation. Prentice–Hall.
brochu, eric, vlad m. cora, and nando de freitas (2010). A Tutorial on Bayesian Opti-
mization of Expensive Cost Functions, with Application to Active User Modeling
and Hierarchical Reinforcement Learning. arXiv: 1012.2599 [cs.LG].
brooks, steve, andrew gelman, galin l. jones, and xiao-li meng, eds. (2011). Handbook
of Markov Chain Monte Carlo. Handbooks of Modern Statistical Methods. Chapman
& Hall.
bubeck, sébastien and nicolò cesa-bianchi (2012). Regret Analysis of Stochastic and
Nonstochastic Multi-armed Bandit Problems. Foundations and Trends in Machine
Learning 5(1):1–122.
bubeck, sébastien, rémi munos, and gilles stoltz (2009). Pure Exploration in Multi-
armed Bandits Problems. Proceedings of the 20th International Conference on Algo-
rithmic Learning Theory (alt 2009). Vol. 5809. Lecture Notes in Computer Science.
Springer–Verlag, pp. 23–37.
bull, adam d. (2011). Convergence Rates of Efficient Global Optimization Algorithms.
Journal of Machine Learning Research 12(88):2879–2904.
caflisch, russel e. (1998). Monte Carlo and quasi-Monte Carlo methods. Acta Numerica 7:
1–49.

332
 bibliography

cai, xu, selwyn gomes, and jonathan scarlett (2021). Lenient Regret and Good-Action
Identification in Gaussian Process Bandits. Proceedings of the 38th International
Conference on Machine Learning (icml 2021). Vol. 139. Proceedings of Machine
Learning Research, pp. 1183–1192.
cakmak, sait, raul astudillo, peter frazier, and enlu zhou (2020). Bayesian Opti-
mization of Risk Measures. Advances in Neural Information Processing Systems 33
(neurips 2020), pp. 18039–18049.
calandra, roberto, jan peters, carl edward rasmussen, and marc peter deisen-
roth (2016). Manifold Gaussian Processes for Regression. Proceedings of the 2016
International Joint Conference on Neural Networks (ijcnn 2016), pp. 3338–3345.
calvin, j. and a. žilinskas (1999). On the Convergence of the P-Algorithm for One-
Dimensional Glboal Optimization of Smooth Functions. Journal of Optimization
Theory and Applications 102(3):479–495.
calvin, james m. (1993). Consistency of a Myopic Bayesian Algorithm for One-Dimensional
Global Optimization. Journal of Global Optimization 3(2):223–232.
calvin, james m. (2000). Convergence Rate of the P-Algorithm for Optimization of Con-
tinuous Functions. In: Approximation and Complexity in Numerical Optimization:
Continuous and Discrete Problems. Ed. by panos m. pardalos. Vol. 42. Nonconvex
Optimization and Its Applications. Springer–Verlag, pp. 116–129.
calvin, james m. and antanas žilinskas (2001). On Convergence of a P-Algorithm Based
on a Statistical Model of Continuously Differentiable Functions. Journal of Global
Optimization 19(3):229–245.
camilleri, romain, julian katz-samuels, and kevin jamieson (2021). High-Dimensional
Experimental Design and Kernel Bandits. Proceedings of the 38th International
Conference on Machine Learning (icml 2021). Vol. 139. Proceedings of Machine
Learning Research, pp. 1227–1237.
chapelle, olivier and lihong li (2011). An Empirical Evaluation of Thompson Sampling.
Advances in Neural Information Processing Systems 24 (neurips 2011), pp. 2249–2257.
chernoff, herman (1959). Sequential Design of Experiments. The Annals of Mathematical
Statistics 30(3):755–770.
chernoff, herman (1972). Sequential Analysis and Optimal Design. cbms–nsf Regional
Conference Series in Applied Mathematics. Society for Industrial and Applied
Mathematics.
chevalier, clément and david ginsbourger (2013). Fast Computation of the Multi-points
Expected Improvement with Applications in Batch Selection. Proceedings of the 7th
Learning and Intelligent Optimization Conference (lion 7). Vol. 7997. Lecture Notes
in Computer Science. Springer–Verlag, pp. 59–69.
chowdhury, sayak ray and aditya gopalan (2017). On Kernelized Multi-armed Bandits.
Proceedings of the 34th International Conference on Machine Learning (icml 2017).
Vol. 70. Proceedings of Machine Learning Research, pp. 844–853.
clark, charles e. (1961). The Greatest of a Finite Set of Random Variables. Operations
Research 9(2):145–162.
contal, emile, david buffoni, alexandre robicqet, and nicolas vayatis (2013). Paral-
lel Gaussian Process Optimization with Upper Confidence Bound and Pure Explo-
ration. Proceedings of the 2013 European Conference on Machine Learning and Prin-
ciples and Practice of Knowledge Discovery in Databases (ecml pkdd 2013). Vol. 8188.
Lecture Notes in Computer Science. Springer–Verlag, pp. 225–240.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 333

bibliography

cover, thomas m. and joy a. thomas (2006). Elements of Information Theory. 2nd ed. John
Wiley & Sons.
cox, dennis d. and susan john (1992). A Statistical Method for Global Optimization.
Proceedings of the 1992 ieee International Conference on Systems, Man, and Cybernetics
(smc 1992), pp. 1241–1246.
cunningham, john p., philipp hennig, and simon lacoste-julien (2011). Gaussian Prob-
abilities and Expectation Propagation. arXiv: 1111.6832 [stat.ML].
cutajar, kurt, michael a. osborne, john p. cunningham, and maurizio filippone (2016).
Preconditioning Kernel Matrices. Proceedings of the 33rd International Conference on
Machine Learning (icml 2016). Vol. 48. Proceedings of Machine Learning Research,
pp. 2529–2538.
dai, zhongxiang, haibin yu, bryan kian hsiang low, and patrick jaillet (2019).
Bayesian Optimization Meets Bayesian Optimal Stopping. Proceedings of the 36th
International Conference on Machine Learning (icml 2019). Vol. 97. Proceedings of
Machine Learning Research, pp. 1496–1506.
dalibard, valentin, michael schaarschmidt, and eiko yoneki (2017). boat: Build-
ing Auto-Tuners with Structured Bayesian Optimization. Proceedings of the 26th
International Conference on World Wide Web (www 2017), pp. 479–488.
dani, varsha, thomas p. hayes, and sham m. kakade (2008). Stochastic Linear Optimiza-
tion Under Bandit Feedback. Proceedings of the 21st Conference on Learning Theory
(colt 2008), pp. 355–366.
davis, philip j. and philip rabinowitz (1984). Methods of Numerical Integration. 2nd ed.
Computer Science and Applied Mathematics. Academic Press.
de ath, george, richard m. everson, alma a. rahat, and jonathan e. fieldsend (2021).
Greed is Good: Exploration and Exploitation Trade-offs in Bayesian Optimisation.
acm Transactions on Evolutionary Learning and Optimization 1(1):1–22.
de ath, george, jonathan e. fieldsend, and richard m. everson (2020). What do you
Mean? The Role of the Mean Function in Bayesian Optimization. Proceedings of the
2020 Genetic and Evolutionary Computation Conference (gecco 2020), pp. 1623–1631.
de freitas, nando, alex j. smola, and masrour zoghi (2012a). Regret Bounds for Deter-
ministic Gaussian Process Bandits. arXiv: 1203.2177 [cs.LG].
de freitas, nando, alex j. smola, and masrour zohgi (2012b). Exponential Regret Bounds
for Gaussian Process Bandits with Deterministic Observations. Proceedings of the
29th International Conference on Machine Learning (icml 2012), pp. 955–962.
degroot, morris h. (1970). Optimal Statistical Decisions. McGraw–Hill.
desautels, thomas, andreas krause, and joel w. burdick (2014). Parallelizing Exploration–
Exploitation Tradeoffs in Gaussian Process Bandit Optimization. Journal of Machine
Learning Research 15(119):4053–4103.
diaconis, persi (1988). Bayesian Numerical Analysis. In: Statistical Decision Theory and
Related Topics iv. Ed. by shanti s. gupta and james o. berger. Vol. 1, pp. 163–175.
djolonga, josip, andreas krause, and volkan cevher (2013). High-Dimensional Gaussian
Process Bandits. Advances in Neural Information Processing Systems 26 (neurips
2013), pp. 1025–1033.
domhan, tobias, jost tobias springenberg, and frank hutter (2015). Speeding up Au-
tomatic Hyperparameter Optimization of Deep Neural Networks by Extrapolation
of Learning Curves. Proceedings of the 24th International Conference on Artificial
Intelligence (ijcai 2015), pp. 3460–3468.

334
 bibliography

duvenaud, david, james robert lloyd, roger grosse, joshua b. tenenbaum, and zoubin
ghahramani (2013). Structure Discovery in Nonparametric Regression through
Compositional Kernel Search. Proceedings of the 30th International Conference on
Machine Learning (icml 2013). Vol. 28. Proceedings of Machine Learning Research,
pp. 1166–1174.
emmerich, michael and boris naujoks (2004). Metamodel Assisted Multiobjective Opti-
misation Strategies and their Application in Airfoil Design. In: Adaptive Computing
in Design and Manufacture vi. Ed. by i. c. parmee, pp. 249–260.
emmerich, michael t. m., kyriakos c. giannakoglou, and boris naujoks (2006). Single-
and Multiobjective Evolutionary Optimization Assisted by Gaussian Random Field
Metamodels. ieee Transactions on Evolutionary Computation 10(4):421–439.
fernández-delgado, manuel, eva cernadas, senén barro, and dinani amorim (2014).
Do we Need Hundreds of Classifiers to Solve Real World Classification Problems?
Journal of Machine Learning Research 15(90):3133–3181.
feurer, matthias, jost tobias springenberg, and frank hutter (2015). Initializing
Bayesian Hyperparameter Optimization via Meta-Learning. Proceedings of 29th
aaai Conference on Artificial Intelligence (aaai 2015), pp. 1128–1135.
fisher, ronald a. (1935). The Design of Experiments. Oliver and Boyd.
fleischer, m. (2003). The Measure of Pareto Optima: Applications to Multi-objective
Metaheuristics. Proceedings of the 2nd International Conference on Evolutionary
Multi-Criterion Optimization (emo 2003). Vol. 2632. Lecture Notes in Computer
Science. Springer–Verlag, pp. 519–533.
forrester, alexander i. j., andy j. keane, and neil w. bressloff (2006). Design and
Analysis of “Noisy” Computer Experiments. aiaa Journal 44(10):2331–2339.
frazier, peter and warren powell (2007). The Knowledge Gradient Policy for Offline
Learning with Independent Normal Rewards. Proceedings of the 2007 ieee Inter-
national Symposium on Approximate Dynamic Programming and Reinforcement
Learning (adprl 2007), pp. 143–150.
frazier, peter, warren powell, and savas dayanik (2009). The Knowledge-Gradient
Policy for Correlated Normal Beliefs. informs Journal on Computing 21(4):599–613.
friedman, milton and l. j. savage (1947). Planning Experiments Seeking Maxima. In:
Selected Techniques of Statistical Analysis for Scientific and Industrial Research, and
Production and Management Engineering. Ed. by churchill eisenhart, millard w.
hartay, and w. allen wallis. McGraw–Hill, pp. 363–372.
fröhlich, lukas p., edgar d. klenske, julia vinogradska, christian daniel, and
melanie n. zeilinger (2020). Noisy-Input Entropy Search for Efficient Robust
Bayesian Optimization. Proceedings of the 23rd International Conference on Artificial
Intelligence and Statistics (aistats 2020). Vol. 108. Proceedings of Machine Learning
Research, pp. 2262–2272.
garcía-barcos, javier and ruben martinez-cantin (2021). Robust policy search for
robot navigation. ieee Robotics and Automation Letters 6(2):2389–2396.
gardner, jacob r., chuan guo, kilian q. weinberger, roman garnett, and roger
grosse (2017). Discovering and Exploiting Additive Structure for Bayesian Opti-
mization. Proceedings of the 20th International Conference on Artificial Intelligence
and Statistics (aistats 2017). Vol. 54. Proceedings of Machine Learning Research,
pp. 1311–1319.
gardner, jacob r., matt j. kusner, zhixiang (eddie) xu, kilian q. weinberger, and
john p. cunningham (2014). Bayesian Optimization with Inequality Constraints.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 335

bibliography

Proceedings of the 31st International Conference on Machine Learning (icml 2014).

Vol. 32. Proceedings of Machine Learning Research, pp. 937–945.
gardner, jacob r., geoff pleiss, david bindel, kilian q. weinberger, and andrew
gordon wilson (2018). GPyTorch: Blackbox Matrix–Matrix Gaussian Process In-
ference with gpu Acceleration. Advances in Neural Information Processing Systems
31 (neurips 2018), pp. 7576–7586.
garnett, roman, thomas gärtner, martin vogt, and jürgen bajorath (2015). In-
troducing the ‘active search’ method for iterative virtual screening. Journal of
Computer-Aided Molecular Design 29(4):305–314.
garnett, roman, yamuna krishnamurthy, xuehan xiong, jeff schneider, and richard
mann (2012). Bayesian Optimal Active Search and Surveying. Proceedings of the
29th International Conference on Machine Learning (icml 2012), pp. 1239–1246.
garnett, r., m. a. osborne, and s. j. roberts (2010). Bayesian Optimization for Sensor Set
Selection. Proceedings of the 9th acm/ieee International Conference on Information
Processing in Sensor Networks (ipsn 2010), pp. 209–219.
garnett, roman, michael a. osborne, and philipp hennig (2014). Active Learning of
Linear Embeddings for Gaussian Processes. Proceedings of the 30th Conference on
Uncertainty in Artificial Intelligence (uai 2014), pp. 230–239.
gelbart, michael a., jasper snoek, and ryan p. adams (2014). Bayesian Optimization
with Unknown Constraints. Proceedings of the 30th Conference on Uncertainty in
Artificial Intelligence (uai 2014), pp. 250–259.
gelman, andrew and aki vehtari (2021). What are the most important statistical ideas of
the past 50 years? Journal of the American Statistical Association.
gergonne, joseph diez (1815). Application de la méthode des moindres quarrés à l’interpola-
tion des suites. Annales de Mathématiques pures et appliquées 6:242–252.
ghosal, subhashis and anindya roy (2006). Posterior Consistency of Gaussian Process
Prior for Nonparametric Binary Regression. The Annals of Statistics 34(5):2413–2429.
gibbs, mark n. (1997). Bayesian Gaussian Processes for Regression and Classification.
PhD thesis. University of Cambridge.
gilboa, elad, yunus saatçi, and john p. cunningham (2013). Scaling Multidimensional
Gaussian Processes using Projected Additive Approximations. Proceedings of the
30th International Conference on Machine Learning (icml 2013). Vol. 28. Proceedings
of Machine Learning Research, pp. 454–461.
ginsbourger, david and rodolphe le riche (2010). Towards Gaussian Process-based
Optimization with Finite Time Horizon. Proceedings of the 9th International Work-
shop in Model-Oriented Design and Analysis (moda 9). Contributions to Statistics.
Springer–Verlag, pp. 89–96.
ginsbourger, david, rodolphe le riche, and laurent carraro (2010). Kriging is well-
suited to parallelize optimization. In: Computational Intelligence in Expensive Opti-
mization Problems. Ed. by yoel yenne and chi-keong go. Adaptation Learning and
Optimization. Springer–Verlag, pp. 131–162.
golovin, daniel and qiuyi (richard) zhang (2020). Random Hypervolume Scalariza-
tions for Provable Multi-Objective Black Box Optimization. Proceedings of the 37th
International Conference on Machine Learning (icml 2020). Vol. 119. Proceedings of
Machine Learning Research, pp. 11096–11105.
golub, gene h. and charles f. van loan (2013). Matrix Computations. 4th ed. Johns
Hopkins Studies in the Mathematical Sciences. Johns Hopkins University Press.

336
 bibliography

gómez-bombarelli, rafael, jennifer n. wei, david duvenaud, josé miguel hernández-

lobato, benjamín sánchez-lengeling, dnnis sheberla, et al. (2018). Automatic
Chemical Design Using a Data-Driven Continuous Representation of Molecules.
acs Central Science 4(2):268–276.
gonzález, javier, zhenwen dai, philipp hennig, and neil lawrence (2016a). Batch
Bayesian Optimization via Local Penalization. Proceedings of the 19th International
Conference on Artificial Intelligence and Statistics (aistats 2016). Vol. 51. Proceedings
of Machine Learning Research, pp. 648–657.
gonzález, javier, michael osborne, and neil d. lawrence (2016b). glasses: Relieving
The Myopia Of Bayesian Optimization. Proceedings of the 19th International Con-
ference on Artificial Intelligence and Statistics (aistats 2016). Vol. 51. Proceedings of
Machine Learning Research, pp. 790–799.
gramacy, robert b. and herbert k. h. lee (2011). Optimization Under Unknown Con-
straints. In: Bayesian Statistics 9. Ed. by j. m. bernardo, m. j. bayarri, j. o. berger,
a. p. dawid, d. heckerman, a. f. m. smith, et al. Oxford University Press, pp. 229–
256.
granmo, ole-christoffer (2010). Solving two-armed Bernoulli bandit problems using a
Bayesian learning automaton. International Journal of Intelligent Computing and
Cybernetics 3(2):207–234.
grünewälder, steffen, jean-yves audibert, manfred opper, and john shawe-taylor
(2010). Regret Bounds for Gaussian Process Bandit Problems. Proceedings of the
13th International Conference on Artificial Intelligence and Statistics (aistats 2010).
Vol. 9. Proceedings of Machine Learning Research, pp. 273–280.
hansen, nikolaus (2016). The cma Evolution Strategy: A Tutorial. arXiv: 1604.00772
[cs.LG].
hastie, trevor and robert tibshirani (1986). Generalized Additive Models. Statistical
Science 1(3):297–318.
hennig, philipp, michael a. osborne, and mark girolami (2015). Probabilistic numerics
and uncertainty in computations. Proceedings of the Royal Society A: Mathematical,
Physical and Engineering Sciences 471(2179):20150142.
hennig, philipp, michael a. osborne, and hans kersting (2022). Probabilistic Numerics:
Computation as Machine Learning. Cambridge University Press.
hennig, philipp and christian j. schuler (2012). Entropy Search for Information-Efficient
Global Optimization. Journal of Machine Learning Research 13(Jun):1809–1837.
hensman, james, alexander g. de g. matthews, maurizio filippone, and zoubin
ghahramani (2015). mcmc for Variationally Sparse Gaussian Processes. Advances
in Neural Information Processing Systems 28 (neurips 2015), pp. 1648–1656.
hernández-lobato, daniel, josé miguel hernández-lobato, amar shah, and ryan p.
adams (2016a). Predictive Entropy Search for Multi-objective Bayesian Optimization.
Proceedings of the 33rd International Conference on Machine Learning (icml 2016).
Vol. 48. Proceedings of Machine Learning Research, pp. 1492–1501.
hernández-lobato, josé miguel, michael a. gelbart, ryan p. adams, matthew w.
hoffman, and zoubin ghahramani (2016b). A General Framework for Constrained
Bayesian Optimization using Information-based Search. Journal of Machine Learning
Research 17:1–53.
hernández-lobato, josé miguel, matthew w. hoffman, and zoubin ghahramani
(2014). Predictive Entropy Search for Efficient Global Optimization of Black-box

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 337

bibliography

Functions. Advances in Neural Information Processing Systems 27 (neurips 2014),

pp. 918–926.
hernández-lobato, josé miguel, james reqeima, edward o. pyzer-knapp, and alán
aspuru-guzik (2017). Parallel and Distributed Thompson Sampling for Large-scale
Accelerated Exploration of Chemical Space. Proceedings of the 34th International
Conference on Machine Learning (icml 2017). Vol. 70. Proceedings of Machine Learn-
ing Research, pp. 1470–1479.
hestenes, magnus r. and eduard stiefel (1952). Methods of Conjugate Gradients for
Solving Linear Systems. Journal of Research of the National Bureau of Standards
49(6):409–436.
hie, brian l. and kevin k. yang (2021). Adaptive machine learning for protein engineering.
arXiv: 2106.05466 [q-bio.QM].
hoang, trong nghia, qang minh hoang, ruofei ouyang, and kian hsiang low
(2018). Decentralized High-Dimensional Bayesian Optimization with Factor Graphs.
Proceedings of 32nd aaai Conference on Artificial Intelligence (aaai 2018), pp. 3231–
3238.
hoffman, matthew d. and andrew gelman (2014). The No-U-turn Sampler: Adaptively
Setting Path Lengths in Hamiltonian Monte Carlo. Journal of Machine Learning
Research 15(4):1593–1623.
hoffman, matthew w. and zoubin ghahramani (2015). Output-Space Predictive Entropy
Search for Flexible Global Optimization. Bayesian Optimization: Scalability and
Flexibility Workshop (BayesOpt 2015), Conference on Neural Information Processing
Systems (neurips 2015).
hoffman, matthew w. and bobak shahriari (2014). Modular mechanisms for Bayesian
optimization. Bayesian Optimization in Academia and Industry (BayesOpt 2014),
Conference on Neural Information Processing Systems (neurips 2014).
hotelling, harold (1941). Experimental Determination of the Maximum of a Function.
The Annals of Mathematical Statistics 12(1):20–45.
houlsby, neil, josé miguel hernández-lobato, ferenc huszár, and zoubin ghahra-
mani (2012). Collaborative Gaussian Processes for Preference Learning. Advances
in Neural Information Processing Systems 25 (neurips 2012), pp. 2096–2104.
huang, d., t. t. allen, w. i. notz, and r. a. miller (2006a). Sequential kriging optimization
using multiple-fidelity evaluations. Structural and Multidisciplinary Optimization
32(5):369–382.
huang, d., t. t. allen, w. i. notz, and n. zeng (2006b). Global Optimization of Stochastic
Black-Box Systems via Sequential Kriging Meta-Models. Journal of Global Opti-
mization 34(3):441–466.
hutter, frank, holger h. hoos, and kevin leyton-brown (2011). Sequential Model-
Based Optimization for General Algorithm Configuration. Proceedings of the 5th
Learning and Intelligent Optimization Conference (lion 5). Vol. 6683. Lecture Notes
in Computer Science. Springer–Verlag, pp. 507–523.
hutter, frank, lin xu, holger h. hoos, and kevin leyton-brown (2014). Algorithm
runtime prediction: Methods & evaluation. Artificial Intelligence 206:79–111.
ingersoll jr., jonathan e. (1987). Theory of Financial Decision Making. Rowman & Little-
field Studies in Financial Economics. Rowman & Littlefield.
irwin, john j., khanh g. tang, jennifer young, chinzorig dandarchuluun, benjamin
r. wong, munkhzul khurelbaatar, et al. (2020). zinc20 – A Free Ultralarge-Scale

338
 bibliography

Chemical Database for Ligand Discovery. Journal of Chemical Information and

Modeling 60(12):6065–6073.
janz, david, david r. burt, and javier gonzález (2020). Bandit optimisation of functions
in the Matérn kernel rkhs. Proceedings of the 23rd International Conference on
Artificial Intelligence and Statistics (aistats 2020). Vol. 108. Proceedings of Machine
Learning Research, pp. 2486–2495.
jiang, shali, henry chai, javier gonzález, and roman garnett (2020a). binoculars
for Efficient, Nonmyopic Sequential Experimental Design. Proceedings of the 37th
International Conference on Machine Learning (icml 2020). Vol. 119. Proceedings of
Machine Learning Research, pp. 4794–4803.
jiang, shali, daniel r. jiang, maximilian balandat, brian karrer, jacob r. gardner,
and roman garnett (2020b). Efficient Nonmyopic Bayesian Optimization via
One-Shot Multi-Step Trees. Advances in Neural Information Processing Systems 33
(neurips 2020), pp. 18039–18049.
jiang, shali, gustavo malkomes, geoff converse, alyssa shofner, benjamin moseley,
and roman garnett (2017). Efficient Nonmyopic Active Search. Proceedings of the
34th International Conference on Machine Learning (icml 2017). Vol. 70. Proceedings
of Machine Learning Research, pp. 1714–1723.
jones, d. r., c. d. perttunen, and b. e. stuckman (1993). Lipschitzian Optimization Without
the Lipschitz Constant. Journal of Optimization Theory and Application 79(1):157–181.
jones, donald r. (2001). A Taxonomy of Global Optimization Methods Based on Response
Surfaces. Journal of Global Optimization 21(4):345–383.
jones, donald r., matthias schonlau, and william j. welch (1998). Efficient Global
Optimization of Expensive Black-Box Functions. Journal of Global Optimization
13(4):455–492.
jylänki, pasi, jarno vanhatalo, and aki vehtari (2011). Robust Gaussian Process Re-
gression with a Student-𝑡 Likelihood. Journal of Machine Learning Research 12(99):
3227–3257.
kanagawa, motonobu, philipp hennig, dino sejdinovic, and bharath k. sriperum-
budur (2018). Gaussian Processes and Kernel Methods: A Review on Connections
and Equivalences. arXiv: 1807.02582 [stat.ML].
kandasamy, kirthevasan, gautam dasarathy, junier oliva, jeff schneider, and
barnabás póczos (2016). Gaussian Process Bandit Optimisation with Multi-fidelity
Evaluations. Advances in Neural Information Processing Systems 29 (neurips 2016),
pp. 992–1000.
kandasamy, kirthevasan, gautam dasarathy, jeff schneider, and barnabás póc-
zos (2017). Multi-fidelity Bayesian Optimisation with Continuous Approximations.
Proceedings of the 34th International Conference on Machine Learning (icml 2017).
Vol. 70. Proceedings of Machine Learning Research, pp. 1799–1808.
kandasamy, kirthevasan, akshay krishnamurthy, jeff schneider, and barnabás póc-
zos (2018). Parallelised Bayesian Optimisation via Thompson Sampling. Proceedings
of the 21st International Conference on Artificial Intelligence and Statistics (aistats
2018). Vol. 84. Proceedings of Machine Learning Research, pp. 133–142.
kandasamy, kirthevasan, jeff schneider, and barnabás póczos (2015). High Dimen-
sional Bayesian Optimisation and Bandits via Additive Models. Proceedings of the
32nd International Conference on Machine Learning (icml 2015). Vol. 37. Proceedings
of Machine Learning Research, pp. 295–304.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 339

bibliography

kathuria, tarun, amit deshpande, and pushmeet kohli (2016). Batched Gaussian Pro-
cess Bandit Optimization via Determinantal Point Processes. Advances in Neural
Information Processing Systems 29 (neurips 2016), pp. 4206–4214.
kim, jeankyung and david pollard (1990). Cube Root Asymptotics. The Annals of Statistics
18(1):191–219.
kim, jungtaek and seungjin choi (2020). On Local Optimizers of Acquisition Functions
in Bayesian Optimization. Lecture Notes in Computer Science 12458:675–690.
kim, jungtaek, michael mccourt, tackgeun you, saehoon kim, and seungjin choi
(2021). Bayesian optimization with approximate set kernels. Machine Learning 110(5):
857–879.
klein, aaron, simon bartels, stefan falkner, philipp hennig, and frank hutter
(2015). Towards efficient Bayesian Optimization for Big Data. Bayesian Optimiza-
tion: Scalability and Flexibility Workshop (BayesOpt 2015), Conference on Neural
Information Processing Systems (neurips 2015).
klein, aaron, stefan falkner, jost tobias springenberg, and frank hutter (2017).
Learning Curve Prediction with Bayesian Neural Networks. Proceedings of the 5th
International Conference on Learning Representations (iclr 2017).
knowles, joshua (2005). Parego: A Hybrid Algorithm With On-Line Landscape Approxi-
mation for Expensive Multiobjective Optimization Problems. ieee Transactions on
Evolutionary Computation 10(1):50–66.
ko, chun-wa, jon lee, and maurice qeyranne (1995). An Exact Algorithm for Maximum
Entropy Sampling. Operations Research 43(4):684–691.
konishi, sadanori and genshiro kitagawa (2008). Information Criteria and Statistical
Modeling. Springer Series in Statistics. Springer–Verlag.
kschischang, frank r., brendan j. frey, and hans-andrea leoliger (2001). Factor
Graphs and the Sum–Product Algorithm. ieee Transactions on Information Theory
47(2):498–519.
kulesza, alex and ben taskar (2012). Determinantal Point Processes for Machine Learning.
Foundations and Trends in Machine Learning 5(2–3):123–286.
kushner, harold j. (1962). A Versatile Stochastic Model of a Function of Unknown and
Time Varying Form. Journal of Mathematical Analysis and Applications 5(1):150–167.
kushner, h. j. (1964). A New Method of Locating the Maximum Point of an Arbitrary
Multipeak Curve in the Presence of Noise. Journal of Basic Engineering 86(1):97–106.
kuss, malte (2006). Gaussian Process Models for Robust Regression, Classification, and
Reinforcement Learning. PhD thesis. Technische Universität Darmstadt.
lai, t. l. and herbert robbins (1985). Asymptotically Efficient Adaptive Allocation Rules.
Advances in Applied Mathematics 6(1):4–22.
lam, remi r., karen e. wilcox, and david h. wolpert (2016). Bayesian Optimization with
a Finite Budget: An Approximate Dynamic Programming Approach. Advances in
Neural Information Processing Systems 29 (neurips 2016), pp. 883–891.
lange, kenneth l., roderick j. a. little, and jeremy m. g. taylor (1989). Robust Statistical
Modeling Using the 𝑡 Distribution. Journal of the American Statistical Association
84(408):881–896.
lattimore, tor and csaba szepesvári (2020). Bandit Algorithms. Cambridge University
Press.
lázaro-gredilla, miguel, joaqin qiñonero-candela, carl edward rasmussen, and
aníbal r. figueiras-vidal (2010). Sparse Spectrum Gaussian Process Regression.
Journal of Machine Learning Research 11(Jun):1865–1881.

340
 bibliography

letham, benjamin, brian karrer, guilherme ottoni, and eytan bakshy (2019). Con-
strained Bayesian Optimization with Noisy Experiments. Bayesian Analysis 14(2):
495–519.
levina, elizaveta and peter j. bickel (2004). Maximum Likelihood Estimation of Intrinsic
Dimension. Advances in Neural Information Processing Systems 17 (neurips 2004),
pp. 777–784.
lévy, paul (1948). Processus stochastiques et mouvement brownien. Gauthier–Villars.
li, chun-liang, kirthevasan kandasamy, barnabás póczos, and jeff schneider (2016).
High Dimensional Bayesian Optimization via Restricted Projection Pursuit Models.
Proceedings of the 19th International Conference on Artificial Intelligence and Statistics
(aistats 2016). Vol. 51. Proceedings of Machine Learning Research, pp. 884–892.
li, chunyuan, heerad farkhoor, rosanne liu, and jason yosinski (2018a). Measuring
the Intrinsic Dimension of Objective Landscapes. Proceedings of the 6th International
Conference on Learning Representations (iclr 2018).
li, lisha, kevin jamieson, giulia desalvo, afshin rostamizadeh, and ameet talwalkar
(2018b). Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimiza-
tion. Journal of Machine Learning Research 18(185):1–52.
li, zihan and jonathan scarlett (2021). Gaussian Process Bandit Optimization with Few
Batches. arXiv: 2110.07788 [stat.ML].
lindley, d. v. (1956). On a Measure of the Information Provided by an Experiment. The
Annals of Mathematical Statistics 27(4):986–1005.
lindley, d. v. (1972). Bayesian Statistics, A Review. cbms–nsf Regional Conference Series
in Applied Mathematics. Society for Industrial and Applied Mathematics.
locatelli, m. (1997). Bayesian Algorithms for One-Dimensional Global Optimization.
Journal of Global Optimization 10(1):57–76.
löwner, karl (1934). Über monotone Matrixfunktionen. Mathematische Zeitschrift 38:
177–216.
lukić, milan n. and jay h. beder (2001). Stochastic Processes with Sample Paths in
Reproducing Kernel Hilbert Spaces. Transactions of the American Mathematical
Society 353(10):3945–3969.
lyu, yueming, yuan yuan, and ivor w. tsang (2019). Efficient Batch Black-box Optimiza-
tion with Deterministic Regret Bounds. arXiv: 1905.10041 [cs.LG].
mackay, david j. c. (1998). Introduction to Gaussian Processes. Neural Networks and
Machine Learning. Ed. by christopher m. bishop. Vol. 168. nato asi Series F:
Computer and Systems Sciences. Springer–Verlag.
mackay, david j. c. (2003). Information Theory, Inference, and Learning Algorithms. Cam-
bridge University Press.
malkomes, gustavo and roman garnett (2018). Automating Bayesian optimization
with Bayesian optimization. Advances in Neural Information Processing Systems 31
(neurips 2018), pp. 5984–5994.
malkomes, gustavo, chip schaff, and roman garnett (2016). Bayesian optimization for
automated model selection. Advances in Neural Information Processing Systems 29
(neurips 2016), pp. 2900–2908.
marmin, sébastien, clément chevalier, and david ginsbourger (2015). Differentiating
the Multipoint Expected Improvement for Optimal Batch Design. Proceedings of the
1st International Workshop on Machine Learning, Optimization, and Big Data (mod
2015). Vol. 9432. Lecture Notes in Computer Science. Springer–Verlag, pp. 37–48.
marschak, jacob and roy radner (1972). Economic Theory of Teams. Yale University Press.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 341

bibliography

martinez-cantin, ruben, kevin tee, and michael mccourt (2018). Practical Bayesian
optimization in the presence of outliers. Proceedings of the 21st International Con-
ference on Artificial Intelligence and Statistics (aistats 2018). Vol. 84. Proceedings of
Machine Learning Research, pp. 1722–1731.
massart, pascal (2007). Concentration Inequalities and Model Selection: Ecole d’Eté de
Probabilités de Saint-Flour xxxiii – 2003. Vol. 1896. Lecture Notes in Mathematics.
Springer–Verlag.
mccullagh, p. and j. a. nelder (1989). Generalized Linear Models. 2nd ed. Monographs on
Statistics and Applied Probability. Chapman & Hall.
meinshausen, nicolai (2006). Quantile Regression Forests. Journal of Machine Learning
Research 7(35):983–999.
miettinen, kaisa m. (1998). Nonlinear Multiobjective Optimization. International Series in
Operations Research & Management Science. Kluwer Academic Publishers.
milgrom, paul and ilya segal (2002). Envelope Theorems for Arbitrary Choice Sets.
Econometrica 70(2):583–601.
minka, thomas (2008). ep: A quick reference. url: https://tminka.github.io/papers/
ep/minka-ep-quickref.pdf.
minka, thomas p. (2001). A family of algorithms for approximate Bayesian inference. Ph.D.
thesis. Massachusetts Institute of Technology.
mockus, jonas (1972). Bayesian Methods of Search for an Extremum. Avtomatika i Vychis-
litel’naya Tekhnika (Automatic Control and Computer Sciences) 6(3):53–62.
mockus, jonas (1974). On Bayesian Methods for Seeking the Extremum. Optimization
Techniques ifip Technical Conference. Vol. 27. Lecture Notes in Computer Science.
Springer–Verlag, pp. 400–404.
mockus, jonas (1989). Bayesian Approach to Global Optimization: Theory and Applications.
Mathematics and its Applications. Kluwer Academic Publishers.
mockus, jonas, william eddy, audris mockus, linas mockus, and gintaras reklaitas
(2010). Bayesian Heuristic Approach to Discrete and Global Optimization: Algorithms,
Visualization, Software, and Applications. Nonconvex Optimization and its Applica-
tions. Kluwer Academic Publishers.
mockus, j., v. tiešis, and a. žilinskas (1978). The Application of Bayesian Methods for
Seeking the Extrememum. In: Towards Global Optimization 2. Ed. by l. c. w. dixon
and g. p. szegö. North–Holland, pp. 117–129.
møller, jesper, anne randi syversveen, and rasmus plenge waagepetersen (1998).
Log Gaussian Cox Processes. Scandinavian Journal of Statistics 25(3):451–482.
montgomery, douglas c. (2019). Design and Analysis of Experiments. 10th ed. John Wiley
& Sons.
moore, andrew w. and christopher g. atkeson (1993). Memory-based Reinforcement
Learning: Efficient Computation with Prioritized Sweeping. Advances in Neural
Information Processing Systems 5 (neurips 1992), pp. 263–270.
moriconi, riccardo, marc peter deisenroth, and k. s. sesh kumar (2020). High-
dimensional Bayesian optimization using low-dimensional feature spaces. Machine
Learning 109(9–10):1925–1943.
moss, henry b., daniel beck, javier gonzález, david s. leslie, and paul rayson (2020).
boss: Bayesian Optimization over String Spaces. Advances in Neural Information
Processing Systems 33 (neurips 2020), pp. 15476–15486.
murray, iain (2016). Differentiation of the Cholesky decomposition. arXiv: 1602.07527
[stat.CO].

342
 bibliography

murray, iain, ryan prescott adams, and david j. c. mackay (2010). Elliptical slice
sampling. Proceedings of the 13th International Conference on Artificial Intelligence
and Statistics (aistats 2010). Vol. 9. Proceedings of Machine Learning Research,
pp. 541–548.
mutný, mojmír and andreas krause (2018). Efficient High Dimensional Bayesian Op-
timization with Additivity and Quadrature Fourier Features. Advances in Neural
Information Processing Systems 31 (neurips 2018), pp. 9005–9016.
neal, radford m. (1997). Monte Carlo Implementation of Gaussian Process Models for Bayes-
ian Regression and Classification. Technical report (9702). Department of Statistics,
University of Toronto.
neal, radford m. (1998). Regression and Classification Using Gaussian Process Priors. In:
Bayesian Statistics 6. Ed. by j. m. bernardo, j. o. berger, a. p. dawid, and a. f. m.
smith. Oxford University Press, pp. 475–490.
nguyen, qoc phong, zhongxiang dai, bryan kian hsiang low, and patrick jaillet
(2021a). Optimizing Conditional Value-At-Risk of Black-Box Functions. Advances
in Neural Information Processing Systems 34 (neurips 2021).
nguyen, qoc phong, zhongxiang dai, bryan kian hsiang low, and patrick jaillet
(2021b). Value-at-Risk Optimization with Gaussian Processes. Proceedings of the
38th International Conference on Machine Learning (icml 2021). Vol. 139. Proceedings
of Machine Learning Research, pp. 8063–8072.
nguyen, vu, sunil gupta, santu rana, cheng li, and svetha venkatesh (2017). Regret
for Expected Improvement over the Best-Observed Value and Stopping Condition.
Proceedings of the 9th Asian Conference on Machine Learning (acml 2017). Vol. 77.
Proceedings of Machine Learning Research, pp. 279–294.
nickisch, hannes and carl edward rasmussen (2008). Appproximations for Binary
Gaussian Process Classification. Journal of Machine Learning Research 9(Oct):2035–
2078.
o’hagan, a. (1978). Curve Fitting and Optimal Design for Prediction. Journal of the Royal
Statistical Society Series B (Methodological) 40(1):1–42.
o’hagan, a. (1991). Bayes–Hermite quadrature. Journal of Statistical Planning and Inference
29(3):245–260.
o’hagan, anthony and jonathan forster (2004). Kendall’s Advanced Theory of Statistics.
2nd ed. Vol. 2b: Bayesian Inference. Arnold.
oh, changyong, jakub m. tomczak, efstratios gavves, and max welling (2019). Com-
binatorial Bayesian Optimization using the Graph Cartesian Product. Advances in
Neural Information Processing Systems 32 (neurips 2019), pp. 2914–2924.
øksendal, bernt (2013). Stochastic Differential Equations: An Introduction with Applications.
6th ed. Universitext. Springer–Verlag.
oliveira, rafael, lionel ott, and fabio ramos (2019). Bayesian optimisation under
uncertain inputs. Proceedings of the 22nd International Conference on Artificial
Intelligence and Statistics (aistats 2019). Vol. 89. Proceedings of Machine Learning
Research, pp. 1177–1184.
osborne, michael a., david duvenaud, roman garnett, carl e. rasmussen, stephen j.
roberts, and zoubin ghahramani (2012). Active Learning of Model Evidence
Using Bayesian Quadrature. Advances in Neural Information Processing Systems 25
(neurips 2012), pp. 46–54.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 343

bibliography

osborne, michael a., roman garnett, and stephen j. roberts (2009). Gaussian Processes
for Global Optimization. Proceedings of the 3rd Learning and Intelligent Optimization
Conference (lion 3).
paria, biswajit, kirthevasan kandasamy, and barnabás póczos (2019). A Flexible
Framework for Multi-Objective Bayesian Optimization using Random Scalarizations.
Proceedings of the 35th Conference on Uncertainty in Artificial Intelligence (uai 2019).
Vol. 115. Proceedings of Machine Learning Research, pp. 766–776.
peirce, c. s. (1876). Note on the Theory of the Economy of Research. In: Report of the
Superintendent of the United States Coast Survey Showing the Progress of the Work
for the Fiscal Year Ending with June, 1876. Government Printing Office, pp. 197–201.
picheny, victor (2014). A Stepwise uncertainty reduction approach to constrained global
optimization. Proceedings of the 17th International Conference on Artificial Intelligence
and Statistics (aistats 2014). Vol. 33. Proceedings of Machine Learning Research,
pp. 787–795.
picheny, victor (2015). Multiobjective optimization using Gaussian process emulators via
stepwise uncertainty reduction. Statistics and Computing 25(6):1265–1280.
picheny, victor, david ginsbourger, yann richet, and gregory caplin (2013a). Quantile-
Based Optimization of Noisy Computer Experiments With Tunable Precision. Tech-
nometrics 55(1):2–13.
picheny, victor, tobias wagner, and david ginsbourger (2013b). A benchmark of
kriging-based infill criteria for noisy optimization. Structural and Multidisciplinary
Optimization 48(3):607–626.
pleiss, geoff, martin jankowiak, david eriksson, anil damle, and jacob r. gardner
(2020). Fast Matrix Square Roots with Applications to Gaussian Processes and
Bayesian Optimization. Advances in Neural Information Processing Systems 33
(neurips 2020), pp. 22268–22281.
poincaré, henri (1912). Calcul des probabilités. 2nd ed. Gauthier–Villars.
ponweiser, wolfgang, tobias wagner, dirk biermann, and markus vincze (2008).
Multiobjective Optimization on a Limited Budget of Evaluations Using Model-
Assisted S-Metric Selection. Proceedings of the 10th International Confernce on
Parallel Problem Solving from Nature (ppsn x). Vol. 5199. Lecture Notes in Computer
Science. Springer–Verlag, pp. 784–794.
powell, warren b. (2011). Approximate Dynamic Programming: Solving the Curses of
Dimensionality. 2nd ed. Wiley Series in Probability and Statistics. John Wiley &
Sons.
pronzato, luc (2017). Minimax and maximin space-filling designs: some properties and
methods for construction. Journal de la Société Française de Statistique 158(1):7–36.
raiffa, howard and robert schlaifer (1961). Applied Statistical Decision Theory. Division
of Research, Graduate School of Business Administration, Harvard University.
rasmussen, carl edward and zoubin ghahramani (2002). Bayesian Monte Carlo. Ad-
vances in Neural Information Processing Systems 15 (neurips 2002), pp. 505–512.
rasmussen, carl edward and christopher k. i. williams (2006). Gaussian Processes for
Machine Learning. Adaptive Computation and Machine Learning. mit Press.
riqelme, carlos, george tucker, and jasper snoek (2018). Deep Bayesian Bandits Show-
down. Proceedings of the 6th International Conference on Learning Representations
(iclr 2018).
robbins, herbert (1952). Some Aspects of the Sequential Design of Experiments. Bulletin
of the American Mathematical Society 58(5):527–535.

344
 bibliography

rolland, paul, jonathan scarlett, ilija bogunovic, and volkan cevher (2018). High-
Dimensional Bayesian Optimization via Additive Models with Overlapping Groups.
Proceedings of the 21st International Conference on Artificial Intelligence and Statistics
(aistats 2018). Vol. 84. Proceedings of Machine Learning Research, pp. 298–307.
ross, andrew m. (2010). Computing Bounds on the Expected Maximum of Correlated
Normal Variables. Methodology and Computing in Applied Probability 12(1):111–138.
rudin, walter (1976). Principles of Mathematical Analysis. 3rd ed. International Series in
Pure and Applied Mathematics. McGraw–Hill.
rue, håvard, sara martino, and nicolas chopin (2009). Approximate Bayesian inference
for latent Gaussian models by using integrated nested Laplace approximations.
Journal of the Royal Statistical Society Series B (Methodological) 71(2):319–392.
russo, daniel and benjamin van roy (2014). Learning to Optimize via Posterior Sampling.
Mathematics of Operations Research 39(4):1221–1243.
russo, daniel and benjamin van roy (2016). An Information-Theoretic Analysis of Thomp-
son Sampling. Journal of Machine Learning Research 17(68):1–30.
sacks, jerome, william j. welch, toby j. mitchell, and henry p. wynn (1989). Design
and Analysis of Computer Experiments. Statistical Science 4(4):409–435.
salgia, sudeep, sattar vakili, and qing zhao (2020). A Computationally Efficient Ap-
proach to Black-box Optimization using Gaussian Process Models. arXiv: 2010.
13997 [stat.ML].
šaltenis, vydūnas r. (1971). One Method of Multiextremum Optimization. Avtomatika i
Vychislitel’naya Tekhnika (Automatic Control and Computer Sciences) 5(3):33–38.
sanchez, susan m. and paul j. sanchez (2005). Very Large Fractional Factorial and Central
Composite Designs. acm Transactions on Modeling and Computer Simulation 15(4):
362–377.
scarlett, jonathan (2018). Tight Regret Bounds for Bayesian Optimization in One
Dimension. Proceedings of the 35th International Conference on Machine Learning
(icml 2018). Vol. 80. Proceedings of Machine Learning Research, pp. 4500–4508.
scarlett, jonathan, ilija bogunovic, and volkan cevher (2017). Lower Bounds on
Regret for Noisy Gaussian Process Bandit Optimization. Proceedings of the 2017
Conference on Learning Theory (colt 2017). Vol. 65. Proceedings of Machine Learning
Research, pp. 1723–1742.
scarlett, jonathan and volkan cevhar (2021). An Introductory Guide to Fano’s Inequal-
ity with Applications in Statistical Estimation. In: Information-Theoretic Methods
in Data Science. Ed. by miguel r. d. rodrigues and yonina c. eldar. Cambridge
University Press.
schonlau, matthias, william j. welch, and donald r. jones (1998). Global versus Local
Search in Constrained Optimization of Computer Models. In: New Developments
and Applications in Experimental Design. Vol. 34. Lecture Notes – Monograph Series.
Institute of Mathematical Statistics, pp. 11–25.
schonlau, mattihias (1997). Computer Experiments and Global Optimization. Ph.D. thesis.
University of Waterloo.
scott, warren, peter frazier, and warren powell (2011). The Correlated Knowledge
Gradient for Simulation Optimization of Continuous Parameters Using Gaussian
Process Regression. siam Journal on Optimization 21(3):996–1026.
seeger, matthias (2008). Expectation Propagation for Exponential Families. Technical
report. University of California, Berkeley.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 345

bibliography

settles, burr (2012). Active Learning. Synthesis Lectures on Artificial Intelligence and
Machine Learning. Morgan & Claypool.
shah, amar and zoubin ghahramani (2015). Parallel Predictive Entropy Search for Batch
Global Optimization of Expensive Objective Functions. Advances in Neural Infor-
mation Processing Systems 28 (neurips 2015), pp. 3330–3338.
shah, amar, andrew gordon wilson, and zoubin ghahramani (2014). Student-𝑡 Pro-
cesses as Alternatives to Gaussian Processes. Proceedings of the 17th International
Conference on Artificial Intelligence and Statistics (aistats 2014). Vol. 33. Proceedings
of Machine Learning Research, pp. 877–885.
shahriari, bobak, kevin swersky, ziyu wang, ryan p. adams, and nando de freitas
(2016). Taking the Human Out of the Loop: A Review of Bayesian Optimization.
Proceedings of the ieee 104(1):148–175.
shahriari, bobak, ziyu wang, matthew w. hoffman, alexandre bouchard-côté, and
nando de freitas (2014). An Entropy Search Portfolio for Bayesian Optimization.
arXiv: 1406.4625 [stat.ML].
shamir, ohad (2013). On the Complexity of Bandit and Derivative-Free Stochastic Convex
Optimization. Proceedings of the 24th Annual Conference on Learning Theory (colt
2013). Vol. 30. Proceedings of Machine Learning Research, pp. 3–24.
shannon, c. e. (1948). A Mathematical Theory of Communication. The Bell System Technical
Journal 27(3):379–423.
shao, t. s., t. c. chen, and r. m. frank (1964). Tables of Zeros and Gaussian Weights of
Certain Associated Laguerre Polynomials and the Related Generalized Hermite
Polynomials. Mathematics of Computation 18(88):598–616.
shepp, l. a. (1979). The Joint Density of the Maximum and its Location for a Wiener Process
with Drift. Journal of Applied Probability 16(2):423–427.
silverman, b. w. (1986). Density Estimation for Statistics and Data Analysis. Monographs
on Statistics and Applied Probability. Chapman & Hall.
slepian, david (1962). The One-Sided Barrier Problem for Gaussian Noise. The Bell System
Technical Journal 41(2):463–501.
smith, kirstine (1918). On the Standard Deviations of Adjusted and Interpolated Values of
an Observed Polynomial Function and its Constants and the Guidance they Give
Towards a Proper Choice of the Distribution of Observations. Biometrika 12(1–2):
1–85.
smola, alex j. and bernhard schölkopf (2000). Sparse Greedy Matrix Approximation
for Machine Learning. Proceedings of the 17th International Conference on Machine
Learning (icml 2000), pp. 911–918.
snelson, edward and zoubin ghahramani (2005). Sparse Gaussian Processes using
Pseudo-inputs. Advances in Neural Information Processing Systems 18 (neurips 2005),
pp. 1257–1264.
snoek, jasper, hugo larochelle, and ryan p. adams (2012). Practical Bayesian Optimiza-
tion of Machine Learning Algorithms. Advances in Neural Information Processing
Systems 25 (neurips 2012), pp. 2951–2959.
snoek, jasper, oren rippel, kevin swersky, ryan kiros, nadathur satish, narayanan
sundaram, et al. (2015). Scalable Bayesian Optimization Using Deep Neural Net-
works. Proceedings of the 32nd International Conference on Machine Learning (icml
2015). Vol. 37. Proceedings of Machine Learning Research, pp. 2171–2180.
snoek, jasper, kevin swersky, richard zemel, and ryan p. adams (2014). Input Warping
for Bayesian Optimization of Non-Sationary Functions. Proceedings of the 31st

346
 bibliography

International Conference on Machine Learning (icml 2014). Vol. 32. Proceedings of

Machine Learning Research, pp. 1674–1682.
song, jialin, yuxin chen, and yisong yue (2019). A General Framework for Multi-fidelity
Bayesian Optimization with Gaussian Processes. Proceedings of the 22nd Inter-
national Conference on Artificial Intelligence and Statistics (aistats 2019). Vol. 89.
Proceedings of Machine Learning Research, pp. 3158–3167.
springenberg, jost tobias, aaron klein, stefan falkner, and frank hutter (2016).
Bayesian Optimization with Robust Bayesian Neural Networks. Advances in Neural
Information Processing Systems 29 (neurips 2016), pp. 4134–4142.
srinivas, niranjan, andreas krause, sham kakade, and matthias seeger (2010). Gaus-
sian Process Optimization in the Bandit Setting: No Regret and Experimental Design.
Proceedings of the 27th International Conference on Machine Learning (icml 2010),
pp. 1015–1022.
stein, michael l. (1999). Interpolation of Spatial Data: Some Theory for Kriging. Springer
Series in Statistics. Springer–Verlag.
streltsov, simon and pirooz vakili (1999). A Non-myopic Utility Function for Statistical
Global Optimization Algorithms. Journal of Global Optimization 14(3):283–298.
sutton, richard s. (1990). Integrated Architectures for Learning, Planning, and Reacting
Based on Approximating Dynamic Programming. Proceedings of the 7th Interna-
tional Conference on Machine Learning (icml 1990), pp. 216–224.
svenson, joshua and thomas santner (2016). Multiobjective optimization of expensive-
to-evaluate deterministic computer simulator models. Computational Statistics and
Data Analysis 94:250–264.
swersky, kevin, jasper snoek, and ryan p. adams (2013). Multi-Task Bayesian Optimiza-
tion. Advances in Neural Information Processing Systems 26 (neurips 2013), pp. 2004–
2012.
swersky, kevin, jasper snoek, and ryan p. adams (2014). Freeze–Thaw Bayesian Opti-
mization. arXiv: 1406.3896 [stat.ML].
takeno, shion, hitoshi fukuoka, yuhki tsukada, toshiyuki koyama, motoki shiga,
ichiro takeuchi, et al. (2020). Multi-fidelity Bayesian Optimization with Max-
value Entropy Search and its Parallelization. Proceedings of the 37th International
Conference on Machine Learning (icml 2020). Vol. 119. Proceedings of Machine
Learning Research, pp. 9334–9345.
tallis, g. m. (1961). The Moment Generating Function of the Truncated Multi-normal
Distribution. Journal of the Royal Statistical Society Series B (Methodological) 23(1):
223–229.
tesch, matthew, jeff schneider, and howie choset (2013). Expensive Function Opti-
mization with Stochastic Binary Outcomes. Proceedings of the 30th International
Conference on Machine Learning (icml 2013). Vol. 28. Proceedings of Machine Learn-
ing Research, pp. 1283–1291.
the uniprot consortium (2021). UniProt: the universal protein knowledgebase in 2021.
Nucleic Acids Research 49(d1):d480–d489.
thompson, william r. (1933). On the Likelihood that One Unknown Probability Exceeds
Another in View of the Evidence of Two Samples. Biometrika 25(3–4):285–294.
thompson, william r. (1935). On the Theory of Apportionment. American Journal of
Mathematics 57(2):450–456.
tiao, louis c., aaron klein, matthias seeger, edwin v. bonilla, cédric archambeau,
and fabio ramos (2021). bore: Bayesian Optimization by Density-Ratio Estimation.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 347

bibliography

Proceedings of the 38th International Conference on Machine Learning (icml 2021).

Vol. 139. Proceedings of Machine Learning Research, pp. 10289–10300.
titsias, michalis (2009). Variational Learning of Inducing Variables in Sparse Gaussian
Processes. Proceedings of the 12th International Conference on Artificial Intelligence
and Statistics (aistats 2009). Vol. 5. Proceedings of Machine Learning Research,
pp. 567–574.
toscano-palmerin, saul and peter i. frazier (2018). Bayesian Optimization with Expen-
sive Integrands. arXiv: 1803.08661 [stat.ML].
turban, sebastien (2010). Convolution of a truncated normal and a centered normal variable.
Technical report. Columbia University.
turner, ryan, david eriksson, michael mccourt, juha kiili, eero laaksonen, zhen
xu, et al. (2021). Bayesian Optimization is Superior to Random Search for Machine
Learning Hyperparameter Tuning: Analysis of the Black-Box Optimization Chal-
lenge 2020. Proceedings of the Neurips 2020 Competition and Demonstration Track.
Vol. 133. Proceedings of Machine Learning Research, pp. 3–26.
ulrich, kyle, david e. carlson, kafui dzirasa, and lawrence carin (2015). gp Kernels
for Cross-Spectrum Analysis. Advances in Neural Information Processing Systems
28 (neurips 2015), pp. 1999–2007.
vakili, sattar, nacime bouziani, sepehr jalali, alberto bernacchia, and da-shan
shiu (2021a). Optimal Order Simple Regret for Gaussian Process Bandits. Advances
in Neural Information Processing Systems 34 (neurips 2021).
vakili, sattar, kia khezeli, and victor picheny (2021b). On Information Gain and
Regret Bounds in Gaussian Process Bandits. Proceedings of the 24th International
Conference on Artificial Intelligence and Statistics (aistats 2021). Vol. 130. Proceedings
of Machine Learning Research, pp. 82–90.
vakili, sattar, jonathan scarlett, and tara javidi (2021c). Open Problem: Tight Online
Confidence Intervals for rkhs Elements. Proceedings of the 34th Annual Conference
on Learning Theory (colt 2021). Vol. 134. Proceedings of Machine Learning Research,
pp. 4647–4652.
valko, michal, nathan korda, rémi munos, ilias flaounas, and nello cristianini
(2013). Finite-Time Analysis of Kernelised Contextual Bandits. Proceedings of the
29th Conference on Uncertainty in Artificial Intelligence (uai 2013), pp. 654–663.
van de corput, j. g. (1935). Verteilungsfunktionen: Erste Mitteilung. Proceedings of the
Koninklijke Nederlandse Akademie van Wetenschappen 38:813–821.
van der vaart, aad w. and jon a. wellner (1996). Weak Convergence and Empirical
Processes with Applications to Statistics. Springer Series in Statistics. Springer–Verlag.
vanchinathan, hastagiri p., andreas marfurt, charles-antoine robelin, donald
kossmann, and andreas krause (2015). Discovering Valuable Items from Massive
Data. Proceedings of the 21st acm sigkdd International Conference on Knowledge
Discovery and Data Mining (kdd 2015), pp. 1195–1204.
vazqez, emmanuel, julien villemonteix, maryan sidorkiewicz, and éric walter
(2008). Global optimization based on noisy evaluations: an empirical study of two
statistical approaches. Proceedings of the 6th International Conference on Inverse
Problems in Engineering: Theory and Practice (icipe 2008). Vol. 135. Journal of Physics:
Conference Series, paper number 012100.
vehtari, aki and janne ojanen (2012). A survey of Bayesian predictive methods for model
assessment, selection and comparison. Statistics Surveys 6:142–228.

348
 bibliography

villemonteix, julien, emmanuel vazqez, and eric walter (2009). An informational

approach to the global optimization of expensive-to-evaluate functions. Journal of
Global Optimization 44(4):509–534.
vivarelli, francesco and christopher k. i williams (1998). Discovering hidden fea-
tures with Gaussian process regression. Advances in Neural Information Processing
Systems 11 (neurips 1998), pp. 613–619.
von neumann, john and oskar morgenstern (1944). Theory of Games and Economic
Behavior. Princeton University Press.
vondrák, jan (2005). Probabilistic Methods in Combinatorial and Stochastic Optimization.
Ph.D. thesis. Massachusetts Institute of Technology.
wald, a. (1945). Sequential Tests of Statistical Hypotheses. The Annals of Mathematical
Statistics 16(2):117–186.
wald, abraham (1947). Sequential Analysis. Wiley Mathematical Statistics Series. John
Wiley & Sons.
wang, jialei, scott c. clark, eric liu, and peter i. frazier (2020a). Parallel Bayesian
Global Optimization of Expensive Functions. Operations Research 68(6):1850–1865.
wang, zexin, vincent y. f. tan, and jonathan scarlett (2020b). Tight Regret Bounds
for Noisy Optimization of a Brownian Motion. arXiv: 2001.09327 [cs.LG].
wang, zi and stefanie jegelka (2017). Max-value Entropy Search for Efficient Bayesian
Optimization. Proceedings of the 34th International Conference on Machine Learning
(icml 2017). Vol. 70. Proceedings of Machine Learning Research, pp. 3627–3635.
wang, zi, bolei zhou, and stefanie jegelka (2016a). Optimization as Estimation with
Gaussian Processes in Bandit Settings. Proceedings of the 19th International Con-
ference on Artificial Intelligence and Statistics (aistats 2016). Vol. 51. Proceedings of
Machine Learning Research, pp. 1022–1031.
wang, ziyu and nando de freitas (2014). Theoretical Analysis of Bayesian Optimization
with Unknown Gaussian Process Hyper-Parameters. arXiv: 1406.7758 [stat.ML].
wang, ziyu, frank hutter, masrour zoghi, david matheson, and nando de freitas
(2016b). Bayesian Optimization in a Billion Dimensions via Random Embeddings.
Journal of Artificial Intelligence Research 55:361–387.
wendland, holger (2004). Scattered Data Approximation. Cambridge Monographs on
Applied and Computational Mathematics. Cambridge University Press.
whittle, peter (1982). Optimization Over Time: Dynamic Programming and Stochastic
Control. Vol. 1. Wiley Series in Probability and Mathematical Statistics. John Wiley
& Sons.
whittle, p. (1988). Restless Bandits: Activity Allocation in a Changing World. Journal of
Applied Probability 25(a):287–298.
williams, brian j., thomas j. santner, and william i. notz (2000). Sequential Design
of Computer Experiments to Minimize Integrated Reseponse Functions. Statistica
Sinica 10(4):1133–1152.
williams, christopher k. i. and matthias seeger (2000). Using the Nyström Method to
Speed Up Kernel Machines. Advances in Neural Information Processing Systems 13
(neurips 2000), pp. 682–688.
wilson, andrew gordon and ryan prescott adams (2013). Gaussian Process Kernels for
Pattern Discovery and Extrapolation. Proceedings of the 30th International Confer-
ence on Machine Learning (icml 2013). Vol. 28. Proceedings of Machine Learning
Research, pp. 1067–1075.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 349

bibliography

wilson, andrew gordon, zhiting hu, ruslan salakhutdinov, and eric p. xing (2016).
Deep Kernel Learning. Proceedings of the 19th International Conference on Artificial
Intelligence and Statistics (aistats 2016). Vol. 51. Proceedings of Machine Learning
Research, pp. 370–378.
wilson, james t., frank hutter, and marc peter deisenroth (2018). Maximizing acquisi-
tion functions for Bayesian optimization. Advances in Neural Information Processing
Systems 31 (neurips 2018), pp. 9884–9895.
wu, jian and peter i. frazier (2016). The Parallel Knowledge Gradient Method for Batch
Bayesian Optimization. Advances in Neural Information Processing Systems 29
(neurips 2016), pp. 3126–3134.
wu, jian and peter i. frazier (2019). Practical Two-Step Look-Ahead Bayesian Optimiza-
tion. Advances in Neural Information Processing Systems 32 (neurips 2019), pp. 9813–
9823.
wu, jian, matthias poloczek, andrew gordon wilson, and peter i. frazier (2017).
Bayesian Optimization with Gradients. Advances in Neural Information Processing
Systems 30 (neurips 2017), pp. 5267–5278.
wu, jian, saul toscano-palmerin, peter i. frazier, and andrew gordon wilson
(2019). Practical Multi-fidelity Bayesian Optimization for Hyperparameter Tuning.
Proceedings of the 35th Conference on Uncertainty in Artificial Intelligence (uai 2019).
Vol. 115. Proceedings of Machine Learning Research, pp. 788–798.
yang, kaifeng, michael emmerich, andré deutz, and thomas bäck (2019a). Efficient
computation of expected hypervolume improvement using box decomposition
algorithms. Journal of Global Optimization 75(1):3–34.
yang, kaifeng, michael emmerich, andré deutz, and thomas bäck (2019b). Multi-
Objective Bayesian Global Optimization using expected hypervolume improvement
gradient. Swarm and Evolutionary Computation 44:945–956.
yue, xubo and raed al kontar (2020). Why Non-myopic Bayesian Optimization is
Promising and How Far Should We Look-ahead? A Study via Rollout. Proceedings
of the 23rd International Conference on Artificial Intelligence and Statistics (aistats
2020). Vol. 108. Proceedings of Machine Learning Research, pp. 2808–2818.
zhang, weitong, dongruo zhou, lihong li, and qanqan gu (2021). Neural Thompson
Sampling. Proceedings of the 9th International Conference on Learning Representations
(iclr 2021.
zhang, yehong, trong nghia hoang, bryan kian hsiang low, and mohan kankan-
halli (2017). Information-Based Multi-Fidelity Bayesian Optimization. Bayesian
Optimization for Science and Engineering Workshop (BayesOpt 2017), Conference on
Neural Information Processing Systems (neurips 2017).
zhou, dongrou, lihong li, and qanqan gu (2020). Neural Contextual Bandits with
ucb-based Exploration. Proceedings of the 37th International Conference on Machine
Learning (icml 2020). Vol. 119. Proceedings of Machine Learning Research, pp. 11492–
11502.
ziatdinov, maxim a., ayana ghosh, and sergei v. kalinin (2021). Physics makes the
difference: Bayesian optimization and active learning via augmented Gaussian
process. arXiv: 2108.10280 [physics.comp-ph].
zilberstein, schlomo (1996). Using Anytime Algorithms in Intelligent Systems. ai Maga-
zine 17(3):73–83.
žilinskas, antanas g. (1975). Single-Step Bayesian Search Method for an Extremum of
Functions of a Single Variable. Kibernetika (Cybernetics) 11(1):160–166.

350
 bibliography

zitzler, eckart (1999). Evolutionary Algorithms for Multiobjective Optimization: Methods

and Applications. PhD thesis. Eidgenössische Technische Hochschule Zürich.
zuluaga, marcela, andreas krause, and markus püschel (2016). 𝜀-pal: An Active Learn-
ing Approach to the Multi-Objective Optimization Problem. Journal of Machine
Learning Research 17(104):1–32.

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 351

INDEX

Primary references and definitions are indicated in bold.

A dynamic termination, 103

a posteriori methods, 270, 272, 274 fixed budget, 91
a priori methods, 270, 273 multi-armed bandits, 143
acquisition function, 𝛼 (𝑥; D), 11, 88, 94, 96, Bayesian inference
98, 150, 247, see also expected introduction to, 6
improvement; knowledge of objective function, 8
gradient; mutual information; Bayesian information criterion, 79, 83
probability of improvement; Bayesian neural networks, 198, 281
upper confidence bound Bayesian Occam’s razor, 71
batch, 𝛽 (x; D), 252 Bayesian quadrature, 33, 278
gradient, 158, 208 Bayesian regret, 𝔼[𝑟𝜏 ], 𝔼[𝑅𝜏 ], 218, 224, 240
optimization, 207 Bellman equation, see Bellman optimality
action space, A, 90, 245 Bellman optimality, 94, 99
for batch observations, 252 beta warping, 58
for dynamic termination, 104 binoculars algorithm, 153
for multifidelity optimization, 263 biology, applications in, 318
for optimization with fixed budget, 91 biomedical engineering, applications in, 320
for sequential procedures, 281 Bochner’s theorem, 50
for terminal recommendation, 110, 250 branch and bound, 232, 240
active learning, 136, 275
active search, 115, 283 C
adaptivity gap, 254 central composite design, 76, 81
additive decomposition, 63, 83 certainty equivalent, 111
aerospace engineering, applications in, 325 characteristic length scale, see length scale
algorithm configuration, 246, 327 chemistry, applications in, 288, 313
anytime algorithm, 209, 239, 248 chemoinformatics, 314
approximate dynamic programming, 100, Cholesky decomposition, 201, 258
150, 284 low-rank updates, 202
augmented Chebyshev scalarization, 274 civil engineering, applications in, 323
automated machine learning (automl), 263, combinatorial optimization, 209
328, see also hyperparameter compactness of domain, 34
tuning conditional entropy, 𝐻 [𝜔 | D], 115, see also
automatic relevance determination (ard), 57, entropy
62, 241 conditional mutual information, 𝐼 (𝜔;𝜓 | D),
automobile engineering, applications in, 326 137, see also mutual information
confidence interval, 225, 233, 237
B conformational search, 315
bandits, see multi-armed bandits conjugate gradients, 203
batch observations, 252, 262 constant liar heuristic, 256
connection to sequential observations, constrained optimization, 249
254 constraint functions, 249
batch rollout, 103, 152, 154, 284, see also unknown, 249
rollout constraints, 249
Bayes’ theorem, 7 constraints on objective function, 36, 39, 56
Bayesian decision theory, 10, 89, 124 continuity in mean square, 29, see also
isolated decisions, 90 sample path continuity
sequential decisions continuous differentiability, 31, 221

This material will be published by Cambridge University Press as Bayesian Optimization. This prepublication 353
version is free to view and download for personal use only. Not for redistribution, resale, or use in derivative
works. ©Roman Garnett 2022.
index

cost-aware optimization, 103, 245, 253, 265 computation without noise, 159
covariance function, see prior covariance gradient, 159
function convergence, 217
cross-covariance function, 19, 23, 24, 27, 30, modified, 154
201, 264 origin, 289
cumulative regret, 𝑅𝜏 , 145, 214 worst-case regret with noise, 239, 243
cumulative reward, 114, 142, 155, 215, 283 expected utility, 90, 93
curse of dimensionality, 61, 208 exploration bonus, 145
exploration vs. exploitation dilemma, 11, 83,
D 123, 128, 131, 133, 143, 145, 146, 148,
de novo design, 314 154, 159, 214, 293
decoupled constraint observations, 252 exponential covariance function, 𝐾M1/2 , 52,
deep kernel, 59, 61 219
deep neural networks, ix, 1, 59, 61, 292 extreme value theorem, 34
density ratio estimation, 197
design and analysis of computer experiments F
(dace), 290 factor graph, 37
determinantal point process, 261, 285 Fano’s inequality, 231
differentiability in mean square, 30, see also feasible region, F, 249
continuous differentiability figure of merit, see acquisition function
differential entropy, 𝐻 [𝜔], see entropy fill distance, 𝛿 x , 238
dilation, 56 Fourier transform, 50, 53, 236
disjoint union, 27 freeze–thaw Bayesian optimization, 281
drug discovery, see molecular design fully independent training conditional (fitc)
dynamic termination, see termination approximation, 207
decisions
G
E Gauss–Hermite quadrature, 172, 258
early stopping, 210, 281 gradient, 309
electrical engineering, applications in, 324 Gaussian process (gp), GP (𝑓 ; 𝜇, 𝐾), 8, 16, 95,
elliptical slice sampling, 38 124, see also prior mean function;
embedding, see linear embedding, neural prior covariance function
embedding approximate inference, 35, see also
entropy search, see mutual information sparse spectrum approximation;
entropy, 𝐻 [𝜔], 115, see also conditional sparse approximation
entropy classification, 41, 283
environmental variables, 277 computation of policies with, 157
expectation propagation, 39, 182, 190, 273, continuity, 28
302 credible intervals, 18
expected gain per unit cost, 248 differentiability, 30
expected hypervolume improvement (ehvi), exact inference, 19
271 additive Gaussian noise, 20, 23
expected improvement, 𝛼 ei , 81, 95, 113, 117, computation, 201
127, 151, 158, 193, 196, 199, 265, 266, derivative observations, 32
268, see also simple reward exact observation, 22
augmented, 166 interpretation of posterior moments,
batch, 259 21
comparison with probability of joint, see joint Gaussian process
improvement, 132 marginal likelihood 𝑝 (y | x, 𝜽 ), 72, 202,
computation with noise, 160 220
alternative formulations, 163 gradient, 307
gradient, 308 maxima, existence and uniqueness, 33

354
 index

mixture, 38, 75, 193 J

model assessment, selection, and joint Gaussian process, 27, 267, see also
averaging, 67 cross-covariance function
modeling, 45 between function and gradient, 30
posterior predictive distribution, 25, 157, exact inference, 28
208 for multifidelity modeling, 264
gradient, 307 marginals, 27
sampling, 18
gene design, 319 K
generalized additive models, 63 kernel, see prior covariance function
generalized linear models, 41 kernel kernel, 83
glasses algorithm, 152 knowledge gradient, 𝛼 kg , 113, 129, 172, 193,
global reward, 113, 117, 129, 172, see also 266, 276, see also global reward
knowledge gradient batch, 259
gp-select algorithm, 285 computation, 172
Gram matrix, 𝐾 (x, x), 17, 49 discrete domain, 173
grid search, 3, 240 gradient, 310
kgcp approximation, 175
H origin, 290
Hamiltonian Monte Carlo (hmc), 75 Kolmogorov extension theorem, 16
Heine–Borel theorem, 30 kriging believer heuristic, 256
heteroskedastic noise, 4, 23, 166 Kullback–Leibler divergence, 𝐷 kl [𝑝 k 𝑞], 115,
Hölder continuity, 35, 221, 240 137, 206
horizon, 93, 125, 151, 245, 254
hubris of youth, ix L
human–computer interfaces, 329 Laplace approximation, 39, 41, 76, 79, 83, 301
hyperband, 282 learning curve, 281
hyperparameter tuning, 1, 61, 109, 263, 267, length scale, 54, 56, 69, see also automatic
291, 327 relevance determination
hyperparameters, 𝜽 , 68, see also length scale; likelihood, 7, see also observation model
output scale marginal, see marginal likelihood
unknown, effect of convergence, 241 limited lookahead, see lookahead
linear covariance function, 𝐾lin , 54
I linear embedding, 58, 62, 209
ill-conditioning, 203 linear scalarization, 274
incumbent value, 𝜙 ∗, 128, 159, 251 linear transformations
inducing values, 𝝊, 205 of domain, 55, 56, 62
infill function, see acquisition function of Gaussian processes, 33, see also
information capacity, 𝛾𝜏 , 221, 225, 229, 233, Bayesian quadrature; joint
237 Gaussian process: between
bounds, 222 function and gradient
information gain, 115, 117, 135, 180, 187, see Lipschitz continuity, 227, 240, 257
also mutual information local optimum, conditioning a Gaussian
initialization, 210 process on a, 36, 182
instantaneous regret, 𝜌𝜏 , 214, 241 local penalization, 257
integrated expected conditional lookahead, 101, 125, 150, 284, see also one-step
improvement, 251 lookahead; two-step lookahead
intrinsic dimensionality, 61 low-dimensional structure, 58, 61, 62, 209,
inverse problems, 317, 319 294
isotropy, 50, 54 low-discrepancy sequence, 177, 211, 239
iterative numerical methods, 203 lower regret bounds
Iverson bracket, xiii Bayesian regret with noise, 230

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 355

index

Bayesian regret without noise, 240 origin, 292

worst-case regret with noise, 236 mutual information, 𝐼 (𝜔;𝜓 ), 116, 135, 291, see
worst-case regret without noise, 239 also information gain; conditional
Löwner order, 22, 48 mutual information
with 𝑓 ,∗ 𝛼 𝑓 ∗ , 140, 187, 193, 266, 291, see
M also output-space predictive
Mahalanobis distance, 48, 57, 297 entropy search; max-value
manifold Gaussian process, 59, 61 entropy search
marginal likelihood, 𝑝 (y | x, 𝜽 ), 71 with 𝑥 ,∗ 𝛼𝑥 ∗ , 139, 180, 193, 266, 291, see
materials science, applications in, 263, 313 also predictive entropy search
Matérn covariance function, 𝐾m , 51, 124, 221, myopic approximation, see lookahead
238, 241
matrix calculus convention, xiii N
maximum a posteriori inference, see model nats, xiii
selection needle in a haystack analogy, 216, 236, 239,
maximum information gain, see information 240
capacity neural architecture search, 328
max-value entropy search, 𝛼 mes , 187 neural embedding, 59, 61, 209
for multiobjective optimization, 273 neural networks, 198
gradient, 191 no u-turn sampler (nuts), 76
mean function, see prior mean function nonmyopic policies, 150, 284, 294
mechanical engineering, applications in, 325 no-regret property, 145, 215
model assessment, 45, 67, 70, see also model Nyström method, 207
posterior
model averaging, 74, 116 O
in acquisition function, 192 objective function posterior, 𝑝 (𝑓 | D), 9, 74,
multiple model structures, 79 92, see also Gaussian process:
model evidence, see marginal likelihood exact inference, approximate
model posterior, 𝑝 (𝜽 | D), 71 inference
model prior, 𝑝 (𝜽 ), 70 model-margninal, see model averaging
model selection, 73, 329 objective function prior, 𝑝 (𝑓 ), 8, see also
multiple model structures, 79 Gaussian process
model space, see model structure observation costs, 104
model structure, M, 68, see also unknown, 245
hyperparameters observation model, 𝑝 (𝑦 | 𝑥, 𝜙), 4
posterior, Pr(M | D), 79 additive Gaussian noise, 4, 23, 69, 78,
prior, Pr(M), 78 157
search, 81 additive Student-𝑡 noise, 36, 38, 41, 282
model, 𝑝 (y | x), 68 exact observation, 4, 22
molecular design, 62, 209, 283, 313 for unknown costs, 246
molecular fingerprint, 314 observation noise scale, 𝜎𝑛 , 4, 23, 69, 78, 157,
Monte Carlo sampling, 37, 75, 84, 181, 187, 258, 203, 222
273 one-step lookahead, 94, 102, 126, 171, 245, 247,
multi-armed bandits, 141 252, 283, 289
infinite-armed, 144 cost-aware optimization, 106
optimal policy, 143 with cumulative reward, 155
multifidelity optimization, 263, 321 with global reward, see knowledge
multiobjective optimization, 269 gradient
multitask optimization, 266 with information gain, see mutual
mutlifidelity optimization, 26 information
mutliobjective optimization, 26 with simple reward, see expected
mutlit-armed bandits improvement

356
 index

one-step lookehead, 251 prior mean function, 𝜇 (𝑥), 16, 46, 67

optimal design, 287 concave quadratic, 48
optimal policy constant, 47, 69, 124, 207
batch observations, 253 marginalization of parameter, 47, 69
computational cost, 99, 125, 284 impact on posterior mean, 46
generic, 245 linear combination of bases, 48
multi-armed bandits, 143 probability of improvement, 𝛼 pi , 131, 167, 193,
sequential optimization, 98 196, 199
optimism in the face of uncertainty, 145 batch, 260
optimization policy, 3, see also acquisition comparison with expected
function; grid search; optimal improvement, 132
policy; random search; Thompson computation with noise, 169
sampling gradient, 309
optimal, see optimal policy computation without noise, 167
Ornstein–Uhlenbeck (ou) process, 52, 174, gradient, 169
290 convergence, 217
output scale, 55, 69, 242 correspondence with upper confidence
output-space predictive entropy search, bound, 170
𝛼 opes , 187 origin, 289
gradient, 311 selection of improvement target, 133,
289
P protein design, 319
Parego algorithm, 275 pseudopoints, see inducing values
Pareto dominance, 270
Pareto frontier, 153, 269 Q
Pareto optimality, see Pareto frontier quantile function, 𝑞(𝜋), 145, 165, 170
Parzen estimation, 197
periodic covariance function, 35, 58, 60 R
physics, applications in, 317 random embedding, 63
plant breeding, 1, 320 random forests, 196
posterior distribution, 7 random search, 3
posterior predictive distribution, 𝑝 (𝑦 | 𝑥, D), reaction optimization, 288, 315
8, 25, 74, 92, 157, 208, 283 regret, 213, see also simple regret; cumulative
for multifidelity modeling, 264 regret; Bayesian regret;
for unknown costs, 246 worst-case regret
model-margninal, see model averaging reinforcement learning, applications in, 322
predictive entropy search, 𝛼 pes , 180 representer points, 180, 188
batch, 260 reproducing kernel Hilbert space (rkhs),
for multiobjective optimization, 273 H𝐾 , 219, 224, 232, 242
gradient, 311 rkhs ball, H𝐾 [𝐵], 220, 224, 232
preference optimization, 282, 330 rkhs norm, k 𝑓 k H𝐾 , 220
prior covariance function, 𝐾 (𝑥, 𝑥 0 ), 17, 49, 67, risk neutrality, 111, 127, 250
see also exponential, linear, risk tolerance, 111
Matérn, spectral mixture, squared risk vs. reward tradeoff, 112, 269
exponential covariance functions; robotics, applications in, 277, 282, 321
automatic relevance robust optimization, 277, 322
determination rolling horizon, 101
addition, 60, 63 rollout, 102, 151, 263
multiplication, 60 batch, see batch rollout
scaling, 55
warping, 56 S
prior distribution, 6 S metric, 271

Draft as of 27 January 2022. Feedback welcome: https://bayesoptbook.com/ 357

index

safe optimization, 322 Bayesian regret with noise, 229

sample path continuity, 30, 34, 218, 221 worst-case regret with noise, 233
scalarization, 273 truncated normal distribution,
second-derivative test, 36, 40 TN (𝜙; 𝜇, 𝜎 ,2 𝐼 ), 40, 159, 189, 305
separable covariance function, 264 two-step lookahead, 96, 150, 251
sequential analysis, 287
sequential experimental design, 288 U
sequential simulation, 255 upper confidence bound, 𝛼 ucb , 145, 170, 195,
signal variance, see output scale 266
simple regret, 𝑟𝜏 , 214, 215, 216, 230, 237 batch, 261
simple reward, 95, 112, 117, 127, 158, 165, 251, computation, 170
284, see also expected correspondence with probability of
improvement improvement, 170
small data, 68 gradient, 170
sparse approximation, 204 origin, 289
sparse spectrum approximation, 51, 178 regret bounds
spectral density, 𝜅, 51, 53 Bayesian regret with noise, 225
spectral measure, 𝜈, 51, 53 worst-case regret with noise, 233,
spectral mixture covariance function, 𝐾sm , 51, 243
53 selecting confidence parameter, 147
spectral points, 178 utility function
sphere packing, 238 for active search, 283
squared exponential covariance function, for constrained optimization, 250
𝐾se , 17, 52, 221 for cost-aware optimization, 104
stationarity, 50, 56, 58, 178, 207, 236 for isolated decisions, 𝑢 (𝑎,𝜓, D), 90
for multifidelity optimization, 264
stochastic optimization, 277
for multitask optimization, 268
stochastic process, 8, 16, see also Gaussian
for optimization, 𝑢 (D), 93, 109, 245, see
process
also cumulative reward; global
structural search, 316
reward; simple reward;
Student-𝑡 process, 282
information gain
sub-Gaussian distribution, 233
for terminal recommendations, 𝜐 (𝜙), 111

T V
terminal recommendation, 118 value of data, 𝛼𝜏∗ , 95, 101
terminal recommendations, 90, 109 value of sample information, 126
termination decisions, 5, 253 variational inference, 39, 206
optimal, 103 virtual screening, 314
practical, 211 von Neumann–Morgenstern theorem, 90, 120
termination option, ∅, 104
Thompson sampling, 148, 176, 181, 187, 195, W
259 weighted Euclidean distance, 57, see also
acquisition function view, 148 automatic relevance
batch, 261 determination
computation, 176 Wiener process, 174, 217, 232, 242, 290
origin, 292 wiggliness, 56, 198, see also length scale
regret bounds worst-case regret, 𝑟¯𝜏 , 𝑅¯𝜏 , 218, 232, 239

358