Distrbutionally Robust Optimization:

A Marriage of Robust Optimization and

Stochastic Programming

Melvyn Sim

Decision Sciences, NUS Business School

Agenda
! Modeling Uncertainty
! DRO
! DRO with Recourse
! Conclusions

3rd Nordic Optimization Symposium

References for this talk
! Chen, W. and M. Sim (2008): Goal Driven Optimization, forthcoming
in Operations Research
! Chen, W., M. Sim, J. Sun and CP Teo (2008): From CVaR to
Uncertainty Sets: Implications in Joint Chance Constrained
Optimization, forthcoming in Operations Research
! Chen, Xin, M. Sim and P. Sun (2007): A Robust Optimization
Perspective of Stochastic Programming, Operations Research, 344-
35755(6), 1058-1071
! Chen, Xin, M. Sim, P. Sun, and J. Zhang (2008): A Linear Decision
based Approximation Approach to Stochastic Programming,
Operations Research, 56(2), 344-357
References for this talk
! K. Natarajan, M. Sim and J. Uichanco (2008): Tractable Robust
Expected Utility and Risk Models for Portfolio Optimization,
forthcoming Mathematical Finance
! K. Natarajan, D. Pachamanova , M. Sim (2008): Incorporating
Asymmetric Distributional Information in Robust Value-at-Risk
Optimization, Management Science, 54(3), 573-585
! See, CT and M. Sim (2008): Robust Approximation to Multi-Period
Inventory Management, under 3rd revision in Operations Research
Agenda
! Modeling Uncertainty
! Robust Linear Optimization
! Robust Linear Optimization with Recourse
! Conclusions

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Probability Distributions
! Data represented as random variables with
known distributions
" Stochastic/Dynamic Programming approach
" Information required
! Sample space (all possible outcomes, usually
exponential or infinite)
! Distributions (probability of outcome)
" Advantages
! A widely accepted method in math/statistics
! Able to quantify expectations such as evaluating
probability of outcomes

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Probability Distributions
! Decision based on taking expectations

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Probability Distributions
" Disadvantages
! Quantifying expectation is computationally intractable

" Shown by Nemirovski and Shapiro 2004, based on a result

of Khachiyan in computing volume of polytope

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Probability Distributions
! Practically prohibitive to obtained exact distributions
" Absence or limited historical data
" Reliability of historical data in predicting outcomes; non-
stationary
" Difficulty of describing multivariate random variable
! How about using empirical distributions or data driven
approaches?

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Probability Distributions
! A Portfolio Optimization Case Study
" 24 small cap stocks from different industry
categories
" Historical returns from April 17 1998 to June 1,
2006
" Return and Covariance estimated from initial 80%
of the data. Evaluate performance on last 20%.

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Probability Distributions
" Markowitz model

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Probability Distributions
" Optimizing over unreliable historical data can be
catastrophic!!!

Markowitz

First Singapore Conference on Quantitative

Finance
 Modeling Uncertainty:
Uncertainty Set
! Data represented as uncertainty set
" Robust Optimization approach
" Information required
! Convex hull of data realization described in tractable
forms:
" Polyhedral
" Conic quadratic, etc

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Uncertainty Set
! Decision based on worst-case value
over the uncertainty set

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Uncertainty Set
! Ellsberg Paradox
" Box 1: 50 red balls and 50 blue balls
" Box 2: 100 red and blue balls with unknown
proportions

Payoffs: $1,000,000 for choosing a red ball.

Which box will you choose?

First Singapore Conference on Quantitative

Finance
Modeling Uncertainty:
Uncertainty Set
! Ellsberg Paradox
" Box 1: 50 red balls and 50 blue balls
" Box 2: 100 red and blue balls with unknown
proportions

Payoffs: $1,000,000 for choosing a blue ball.

Which box will you choose?

Decision Maker is Ambiguity Averse!!

First Singapore Conference on Quantitative
Finance
Modeling Uncertainty:
Uncertainty Set
" Advantages
! Less information is required
" Convex hull versus Sample space
" Distribution free
! Computational tractable for many important classes of
optimization.
! Quantifiable approximation exists for some hard ones
! Natural way of describing uncertainty in certain
applications
" Engineering applications

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Uncertainty Set
" Disadvantages
! Unable to evaluate expectations including probability
measure
! How do we choose the right uncertainty set?
" Requires domain knowledge

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Descriptive Statistics
! Data represented as random variable over a family of
distributions characterized by its descriptive statistics
" Distributionally Robust Optimization approach
" Information required
! Convex hull of support
! Descriptive statistics: means, standard deviations, directional
deviations, independence etc.

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Descriptive Statistics
" Advantages
! Descriptive statistics can be derived from data
! Solutions are robust to distributional assumptions
! Moderate information needed
! Tractable approximations available
! Compute bounds on expectations
! Far less conservative than worst case

3rd Nordic Optimization Symposium

Modeling Uncertainty:
Descriptive Statistics
! Decision based on worst-case value
expectation over the family of
distributions

First Singapore Conference on Quantitative

Finance
Agenda
! Modeling Uncertainty
! DRO
! DRO with Recourse
! Conclusions

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! A typical linear optimization problem:

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! WLOG, we assume data is affinely
dependent on a set of N primitive
uncertainties:
" Provision for linear correlations among data

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Robust Optimization approach
" Find solutions that remains feasible for all data in
uncertainty sets, a.k.a Robust Counterpart

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Stochastic Programming approach

" Chance constraint (Charnce, Cooper and
Symonds 58)

" Typically intractable problem

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Robust chance constrained problem

" Generally intractable

" Tractable formulation for family of distribution with
infinite support, known mean and covariance,
(Bertsimas and Pospescu, 2004, El Ghaoui et al,
2003)

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Should we choose an uncertainty set large
enough to contain most of the samples?

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Key Idea: Convexification of chance
constrained problem.

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Step utility function is not concave!!

! Consider concave approximation

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Conditional Value-at-Risk (CVaR)

" Popularized by Rockafellar and Uryasev
" Best possible convex approximation of chance
constrained problems. (Foellmer and Scheid 2004,
Nemirovski and Shapiro 2006)

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Upper bounds on E( . )+

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Approximation idea:

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

-Chen, Sim, Sun and Teo (2007)

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! X. Chen, S. and P. Sun, 2006

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Worst case deviations with given support

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Unified bound on E( . )+

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Unified bound on CVaR

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Joint chance constraints

" All constraints must be satisfied with high
probability for all random variables with the same
descriptive statistics.
! Much harder to solve than single chance constraint!!!

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! One idea: Union bound

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

! Union bound
" Bound is good if constraints are independently
distributed
" Bound is weak if constraints are highly
correlated.
" Need to fix !j. Sensible choice !j=!/m
! How to optimize over !j?

3rd Nordic Optimization Symposium

Distributionally Robust Optimization
! Can we do better than union bound?
" Yes!! W. Chen, S., J. Sun and Teo (2007)

3rd Nordic Optimization Symposium

Distributionally Robust Optimization

3rd Nordic Optimization Symposium

 Distributionally Robust Optimization:
Resource Allocation Example
! Network of n cities proximally connected
! First stage:
" Decide amount of resouces to place at each city in
anticipation of uncertain demand
! Second stage:
" Demand is realized
" Resouces can be transhipped to neighboring nodes at
zero cost
! Objective
" Find the minimum cost allocation of resouces that meets
service requirement

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example

x1=80
d1=100
x5=50

x4=10 x7=10

x3=250

x6=10
d4=180 x2=90

x8= 100

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example
Infeasible instance

x1=80
x5=50

x4=10 x7=10

x3=250

x6=10
d4=200 x2=90

x8= 100

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example
! Robust joint chance constrained model

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example

! Need to assume linear decision rule on recourse

variables on transshipment

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example
! Computation example: ! = 0.01

Union Bound
Improved Bound

3rd Nordic Optimization Symposium

Distributionally Robust Optimization:
Resource Allocation Example
! Computation example: ! = 0.01

3rd Nordic Optimization Symposium

Agenda
! Modeling Uncertainty
! DRO
! DRO with Recourse
! Conclusions

3rd Nordic Optimization Symposium

DRO with Recourse

! Consider a two stage optimization problem

First stage Second stage

3rd Nordic Optimization Symposium

DRO with Recourse

! Risk Neutral Objective

" Classical stochastic programming model

" Assumes repeatability of experiments under

identical conditions

3rd Nordic Optimization Symposium

DRO with Recourse

! Even when distributions are known,

computations can be difficult
(Dyer and Stougie, 2005)
" Two period models are #P-hard
" >2 periods models are PSPACE-hard

3rd Nordic Optimization Symposium

DRO with Recourse

! Ambiguity Averse, Risk Neutral model

" Famous example: Worst case Newsvendor of

Scarf.

3rd Nordic Optimization Symposium

DRO with Recourse

" Hard problem as well

! Determine a good upper bound
! How good is the bound?

3rd Nordic Optimization Symposium

DRO with Recourse

! Linear Decision Rule Again!!!

" Appeared in early Stochastic Optimization but was
abandon soon.
! Garstka and Wets, 74
" Resurface in adjustable robust counterpart

3rd Nordic Optimization Symposium

DRO with Recourse

! Final Model: Linear Optimization Problem

3rd Nordic Optimization Symposium

DRO with Recourse

! Issues with linear decision rule

" Can lead to infeasible solution even when the
problem has complete recourse

3rd Nordic Optimization Symposium

DRO with Recourse

“The rationale behind restricting to affine decision rules is

the belief that in actual applications it is better to pose a
modest and achievable goal rather than an ambitious
goal which we do not know how to achieve.”
- Shapiro and Nemirovski 05

! Can we do better than linear decision rule?

3rd Nordic Optimization Symposium

DRO with Recourse

! Exploit problem structure of stocastic optimization

model.
" Focus on recourse matrix Y

3rd Nordic Optimization Symposium

DRO with Recourse

! Deflected linear decision rule (X. Chen, S., P. Sun

and Zhang 2006)

3rd Nordic Optimization Symposium

DRO with Recourse

3rd Nordic Optimization Symposium

DRO with Recourse
! Final Model: SOCP
" Uses bound on E( . )+

3rd Nordic Optimization Symposium

DRO with Recourse

! Primitive uncertainties unfolds in stages

! Scales well with Linear Decsion Rule and Deflected

Linear Decision Rules

3rd Nordic Optimization Symposium

Robust Inventory Control

! Multiperiod Inventory Control Problem

" Ordering decision to meet uncertain demand so
that the cost is minimized
" Periodic review, Finite horizon, backlogging,
exogenous demand, no fixed ordering costs

3rd Nordic Optimization Symposium

Robust Inventory Control
! Sequence of events

t t+1

3rd Nordic Optimization Symposium

Robust Inventory Control
! Inventory dynamics

3rd Nordic Optimization Symposium

Robust Inventory Control
! Costs components

3rd Nordic Optimization Symposium

Robust Inventory Control
! Stochastic Optimization Model

3rd Nordic Optimization Symposium

 Robust Inventory Control
! Characterize optimum policy using Dynamic Programming
" Dependent Demand:

! More realistic representation of demand

! Curse of dimensionality
" Independent Demand:
! State independent base-stock policy is optimal
! Does not imply that it is easy to find the base-stock level!!

3rd Nordic Optimization Symposium

Robust Inventory Control

! DRO Inventory Control Model

3rd Nordic Optimization Symposium

Robust Inventory Control
! Factor Demand Model

3rd Nordic Optimization Symposium

Robust Inventory Control
! Factor Demand Model
" Handle demand correlations
" Can include exogenous factors such as market
factors
" Demand forecast models
! E.g: ARMA process

3rd Nordic Optimization Symposium

Robust Inventory Control
! Static Replenishment Policy (Bertsimas and Thiele)
" Inventory position affine in factors

3rd Nordic Optimization Symposium

Robust Inventory Control
! Linear Replenishment Policy

3rd Nordic Optimization Symposium

Robust Inventory Control
" Inventory position affine in factors

3rd Nordic Optimization Symposium

Robust Inventory Control
! Truncated Linear Replenishment Policy (See
and Sim)

3rd Nordic Optimization Symposium

 Robust Inventory Control
! Require the following bound on expectation:
Robust Inventory Control

3rd Nordic Optimization Symposium

Modeling Software

" Sample code implementing TLDR

initx =0*ones(L,1);
inity = 0;

Ny=[0 1:T];
Nx = [zeros(1,L) 0:T-L-1];
Nxms = [zeros(1,L) 0:T-L-1];

% Demand information
Z.zlow = Range*ones(N,1);
Z.zupp = Range*ones(N,1);
Z.p = .58*Range*ones(N,1);
Z.q = .58*Range*ones(N,1);
Z.sigma =.58*Range*ones(N,1);
Modeling Software
startmodel
x = linearrule(T,N,Nx);
xms = linearrule(T,N,Nxms);
y = linearrule(T+1,N,Ny);

for i=1:T
addconst(xms(i,:) == x(i,:)-S*ldrdata([0 1],N));
end

hbound=0;
sbound=0;
for t=1:T

if L+1<= t
hbound = hbound+ h*meannestedposbound(Z,y(t+1,0:t),-x(L+1:t,0:t),t);
sbound = sbound + b(t)*meannestedposbound(Z,-y(t+1,0:t),xms(L+1:t,0:t),t);
else
hbound = hbound+h*meanpositivebound(Z,y(t+1,:),1,N);
sbound = sbound + b(t)*meanpositivebound(Z,-y(t+1,:),1,N);
end

end
Modeling Software

minimize (sbound+hbound + c*sum(meanpositivebound(Z,x(L+1:T,:),T-L,N)))

addconst(x(1:L,0)==initx);
addconst(y(1,0)==inity);
for i=1:T
addconst(y(i+1,:)==y(i,:)+x(i,:)-ldrdata([0 MeanD(i);(1:N)' zcoef(:,i)],N));
end

m=endmodel;
s = m.solve('CPLEX');
xsol=s.eval(x);
Robust Inventory Control
- Computations

3rd Nordic Optimization Symposium

Robust Inventory Control
- Computations
! Compare with
" State independent based-stock policy
! Ignores dependency of previous demands
! Policy is optimal if " = 0
! Use sampling approximation to determine reorder point
" Myopic Policy
! Ignores future costs

3rd Nordic Optimization Symposium

Robust Inventory Control
- Computations
" Truncated Linear Decision Rule

3rd Nordic Optimization Symposium

Robust Inventory Control
- Computations
" Empirical performance (100,000 samples)

3rd Nordic Optimization Symposium

Conclusions
! Robust optimization is a computationally
attractive approach for addressing data
uncertainty in optimization problems
! Many applications
! Many open issues:
" Quantify level of conservativeness
" Address non affine disturbances
" Address general recourse problems
" Address integral recourse problems

3rd Nordic Optimization Symposium