Chance constrained optimization

• chance constraints and percentile optimization

• chance constraints for log-concave distributions

• convex approximation of chance constraints

sources: Rockafellar & Uryasev, Nemirovsky & Shapiro

EE364A — Chance Constrained Optimization 1

 Chance constraints and percentile optimization

• ‘chance constraints’ (η is ‘confidence level’):

Prob(fi(x, ω) ≤ 0) ≥ η

– convex in some cases (later)

– generally interested in η = 0.9, 0.95, 0.99
– η = 0.999 meaningless (unless you’re sure about the distribution tails)

• percentile optimization (γ is ‘η-percentile’):

minimize γ
subject to Prob(f0(x, ω) ≤ γ) ≥ η

– convex or quasi-convex in some cases (later)

EE364A — Chance Constrained Optimization 2

 Value-at-risk and conditional value-at-risk

• value-at-risk of random variable z, at level η:

VaR(z; η) = inf{γ | Prob(z ≤ γ) ≥ η}

– chance constraint Prob(fi(x, ω) ≤ 0) ≥ η same as

VaR(fi(x, ω); η) ≤ 0

• conditional value-at-risk:

CVaR(z; η) = inf (β + 1/(1 − η) E(z − β)+)

β

– CVaR(z; η) ≥ VaR(z; η) (more on this later)

EE364A — Chance Constrained Optimization 3

 CVaR interpretation
(for continuous distributions)
• in CVaR definition, β ⋆ = VaR(z; η):

d
0= (β + 1/(1 − η) E(z − β)+) = 1 − 1/(1 − η) Prob(z ≥ β)
dβ

so Prob(z ≥ β ⋆) = 1 − η

• conditional tail expectation (or expected shortfall)

E(z|z ≥ β ⋆) = E(β ⋆ + (z − β ⋆)|z ≥ β ⋆)

= β ⋆ + E((z − β ⋆)+)/ Prob(z ≥ β ⋆)
= CVaR(z; η)

EE364A — Chance Constrained Optimization 4

 Chance constraints for log-concave distributions

• suppose
– ω has log-concave density p(ω)
– C = {(x, ω) | f (x, ω) ≤ 0} is convex in (x, ω)

• then Z
Prob(f (x, ω) ≤ 0) = 1((x, ω) ∈ C)p(ω) dω
is log-concave, since integrand is

• so chance constraint Prob(f (x, ω) ≤ 0) ≥ η can be expressed as

convex constraint

log Prob(f (x, ω) ≤ 0) ≥ log η

EE364A — Chance Constrained Optimization 5

 Linear inequality with normally distributed parameter

• consider aT x ≤ b, with a ∼ N (ā, Σ)

• then aT x − b ∼ N (āT x − b, xT Σx)

• hence
T
b − ā x
 
Prob(aT x ≤ b) = Φ √
xT Σx

• and so

Prob(aT x ≤ b) ≥ η ⇐⇒ b − āT x ≥ Φ−1(η)kΣ1/2xk2

a second-order cone constraint for η ≥ 0.5 (i.e., Φ−1(η) ≥ 0)

EE364A — Chance Constrained Optimization 6

 Portfolio optimization example

• x ∈ Rn gives portfolio allocation; xi is (fractional) position in asset i

• x must satisfy 1T x = 1, x ∈ C (convex portfolio constraint set)

• portfolio return (say, in percent) is pT x, where p ∼ N (p̄, Σ)

(a more realistic model is p log-normal)

• maximize expected return subject to limit on probability of loss

EE364A — Chance Constrained Optimization 7

• problem is
maximize E pT x
subject to Prob(pT x ≤ 0) ≤ β
1T x = 1, x ∈ C

• can be expressed as convex problem (provided β ≤ 1/2)

maximize p̄T x
subject to p̄T x ≥ Φ−1(1 − β)kΣ1/2xk2
1T x = 1, x ∈ C

(an SOCP when C is polyhedron)

EE364A — Chance Constrained Optimization 8

 Example

• n = 10 assets, β = 0.05, C = {x | x  −0.1}

• compare
– optimal portfolio
– optimal portfolio w/o loss risk constraint
– uniform portfolio (1/n)1

portfolio E pT x Prob(pT x ≤ 0)
optimal 7.51 5.0%
w/o loss constraint 10.66 20.3%
uniform 3.41 18.9%

EE364A — Chance Constrained Optimization 9

return distributions:
optimal

−20 −15 −10 −5 0 5 10 15 20 25 30

w/o loss constraint

−20 −15 −10 −5 0 5 10 15 20 25 30

uniform

−20 −15 −10 −5 0 5 10 15 20 25 30

EE364A — Chance Constrained Optimization 10

 Convex approximation of chance constraint bound

• assume fi(x, ω) is convex in x

• suppose φ : R → R is nonnegative convex nondecreasing, with φ(0) = 1

• for any αi > 0, φ(z/αi) ≥ 1(z > 0) for all z, so

E φ(fi(x, ω)/αi) ≥ Prob(fi(x, ω) > 0)

• hence (convex) constraint

E φ(fi(x, ω)/αi) ≤ 1 − η

ensures chance constraint Prob(fi(x, ω) ≤ 0) ≥ η holds

• this holds for any αi > 0; we now show how to optimize over αi

EE364A — Chance Constrained Optimization 11

• write constraint as

E αiφ(fi(x, ω)/αi) ≤ αi(1 − η)

– (perspective function) vφ(u/v) is convex in (u, v) for v > 0,

nondecreasing in u
– so composition αiφ(fi(x, ω)/αi) is convex in (x, αi) for αi > 0
– hence constraint above is convex in x and αi
– so we can optimize over x and αi > 0 via convex optimization

• yields a convex stochastic optimization problem that is a conservative

approximation of the chance-constrained problem

• we’ll look at some special cases

EE364A — Chance Constrained Optimization 12

 Markov chance constraint bound

• taking φ(u) = (u + 1)+ gives Markov bound: for any αi > 0,

Prob(fi(x, ω) > 0) ≤ E(fi(x, ω)/αi + 1)+

• convex approximation constraint

E αi(fi(x, ω)/αi + 1)+ ≤ αi(1 − η)

can be written as

E(fi(x, ω) + αi)+ ≤ αi(1 − η)

• we can optimize over x and αi ≥ 0

EE364A — Chance Constrained Optimization 13

 Interpretation via conditional value-at-risk
• write conservative approximation as

E(fi(x, ω) + αi)+
− αi ≤ 0
1−η

• LHS is convex in (x, αi), so minimum over αi,

 
E(fi(x, ω) + αi)+
inf − αi
αi >0 1−η

is convex in x

• this is CVaR(fi(x, ω); η) (can show αi > 0 can be dropped)

• so convex approximation replaces VaR(fi(x, ω); η) ≤ 0 with

CVaR(fi(x, ω); η) ≤ 0 which is convex in x

EE364A — Chance Constrained Optimization 14

 Chebyshev chance constraint bound

• taking φ(u) = (u + 1)2+ yields Chebyshev bound: for any αi > 0,

Prob(fi(x, ω) > 0) ≤ E(fi(x, ω)/αi + 1)2+

• convex approximation constraint

E αi(fi(x, ω)/αi + 1)2+ ≤ αi(1 − η)

can be written as

E(fi(x, ω) + αi)2+/αi ≤ αi(1 − η)

EE364A — Chance Constrained Optimization 15

 Traditional Chebyshev bound

• dropping subscript + we get more conservative constraint

E αi(fi(x, ω)/αi + 1)2 ≤ αi(1 − η)

which we can write as

2 E fi(x, ω) + (1/αi) E fi(x, ω)2 + αiη ≤ 0

2

1/2
• minimizing over αi gives αi = E fi(x, ω) /η ; yields constraint

2 1/2


E fi(x, ω) + η E fi(x, ω) ≤0

which depends only on first and second moments of fi

EE364A — Chance Constrained Optimization 16

 Example

• fi(x) = aT x − b, where a is random with E a = ā, E aaT = Σ

• traditional Chebyshev approximation of chance constraint is

T 1/2 T T 2 1/2


ā x − b + η x Σx − 2bā x + b ≤0

• can write as second-order cone constraint

āT x − b + η 1/2k(z, y)k2 ≤ 0

1/2 T

1/2
with z = Σ x − bΣ −1/2
ā, y = b 1 − ā Σ −1
ā

• can interpret as certainty-equivalent constraint, with norm term as

‘extra margin’

EE364A — Chance Constrained Optimization 17

 Chernoff chance constraint bound

• taking φ(u) = exp u yields Chernoff bound: for any αi > 0,

Prob(fi(x, ω) > 0) ≤ E exp(fi(x, ω)/αi)

• convex approximation constraint

E αi exp(fi(x, ω)/αi) ≤ αi(1 − η)

can be written as

log E exp(fi(x, ω)/αi) ≤ log(1 − η)

(LHS is cumulant generating function of fi(x, ω), evaluated at 1/αi)

EE364A — Chance Constrained Optimization 18

 Example

• maximize a linear revenue function (say) subject to random linear

constraints holding with probability η:

maximize cT x
subject to Prob(max(Ax − b) ≤ 0) ≥ η

with variable x ∈ Rn; A ∈ Rm×n, b ∈ Rm random (Gaussian)

• Markov/CVaR approximation:

maximize cT x
subject to E(max(Ax − b) + α)+ ≤ α(1 − η)

with variables x ∈ Rn, α ∈ R

EE364A — Chance Constrained Optimization 19

• Chebyshev approximation:

maximize cT x
subject to E(max(Ax − b) + α)2+/α ≤ α(1 − η)

with variables x ∈ Rn, α ∈ R

• optimal values of these approximate problems are lower bounds for

original problem

EE364A — Chance Constrained Optimization 20

• instance with n = 5, m = 10, η = 0.9

• solve approximations with sampling method with N = 1000 training

samples, validate with M = 10000 samples

• compare to solution of deterministic problem

maximize cT x
subject to E Ax ≤ E b

• estimates of Prob(max(Ax − b) ≤ 0) on training/validation data

cT x train validate
Markov 3.60 0.97 0.96
Chebyshev 3.43 0.97 0.96
deterministic 7.98 0.04 0.03

EE364A — Chance Constrained Optimization 21

• PDF of max(Ax − b) for Markov approximation solution

−25 −20 −15 −10 −5 0 5 10 15

EE364A — Chance Constrained Optimization 22