0% found this document useful (0 votes)
5 views27 pages

Mpro 5

Uploaded by

pokadoc289
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
5 views27 pages

Mpro 5

Uploaded by

pokadoc289
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 27

Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Robust Optimization:
Adaptive Case

V. Leclère (ENPC)

December 15, 2023

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 1 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Contents

1 Fourier-Motzkin Elimination

2 Bender’s decomposition for two-stage problems

3 Linear Decision Rules

4 Piecewise-constant Decision Rules

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 1 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Two-stage robust optimization problem

Recall that a generic robust optimization problem reads as follows:

min max J(x, ξ) + Ig (x,ξ)≤0


x∈Rp ξ∈R

This problem is static or one-stage because the decision maker has to choose x before
knowing the realization of ξ.

In the two-stage or adaptive case, the decision maker can choose a recourse decision y
after observing the realization of ξ:

min max min J(x, ξ, y ) + Ig (x,ξ,y )≤0


x∈Rp ξ∈R y ∈Rq

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 2 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Two-stage robust optimization problem

Recall that a generic robust optimization problem reads as follows:

min max J(x, ξ) + Ig (x,ξ)≤0


x∈Rp ξ∈R

This problem is static or one-stage because the decision maker has to choose x before
knowing the realization of ξ.

In the two-stage or adaptive case, the decision maker can choose a recourse decision y
after observing the realization of ξ:

min max min J(x, ξ, y ) + Ig (x,ξ,y )≤0


x∈Rp ξ∈R y ∈Rq

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 2 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Example
Solve the following example by eliminating the recourse variables:
min x
x∈R,y ∈R[0,1]

s.t. x − yξ ≤ −ξ ∀ξ ∈ [0, 1]
1
− x + yξ ≤ ξ + ∀ξ ∈ [0, 1]
2
yξ ≥ 1 ∀ξ ∈ [0, 1]

Thus, the optimal value is 12 , and there are various optimal y .


V. Leclère Robust Optimization: Adaptive Case December 15, 2023 3 / 14
Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Example
Solve the following example by eliminating the recourse variables:
min x
x∈R,y ∈R[0,1]

s.t. x − yξ ≤ −ξ ∀ξ ∈ [0, 1]
1
− x + yξ ≤ ξ + ∀ξ ∈ [0, 1]
2
yξ ≥ 1 ∀ξ ∈ [0, 1]
The constraints are equivalent to:

 yξ ≥ x +ξ ( n o
1 max 1, ξ + 12 ≤ yξ ≤ 1 + ξ
y ≤ x +ξ+ ⇐⇒
 ξ 2
1 ≤ x +ξ+ 1
yξ ≥ 1 2

Thus, the optimal value is 12 , and there are various optimal y .


V. Leclère Robust Optimization: Adaptive Case December 15, 2023 3 / 14
Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

RO is optimal for constraint-wise uncertainty I

min x
x,y (·)
1
s.t. x − y1 (ξ) ≤ ξ1 ∀ξ1 ∈ [0, ]
2
3
− x + y1 (ξ) + y2 (ξ) ≤ ξ2 + 3ξ3 + ∀ξ2 , ξ3 ∈ [0, 1]2
2
3
1 ≤ y1 (ξ) ≤
2
1 ≤ y2 (ξ)

➥ this problem has constraint-wise uncertainty because the ξ1 affect only the first
constraint, and the ξ2 , ξ3 affect only the second constraint.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 4 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

RO is optimal for constraint-wise uncertainty II

( n o n o
max 1, 0.5 + ξ1 ≤ y1 (ξ) ≤ min 1.5, 1 + ξ2 + 3ξ3
1 ≤ y2 (ξ) ≤ 2 + ξ2 + 3ξ3 − y1 (ξ)
➥ ξ1 appears only on the left-hand side of the first constraint, so we can replace it
with its maximum value 0.5;
similarly, ξ2 , ξ3 appear only on the right-hand side of the second constraint, so we can
replace them with their minimum value 0.
➥ The optimal recourse is y1 (ξ) = y2 (ξ) = 1, that is a constant recourse.

More generally, if the uncertainty is constraint-wise, then the optimal recourse is


constant.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 5 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

RO is optimal for constraint-wise uncertainty II

( n o n o
max 1, 0.5 + ξ1 ≤ y1 (ξ) ≤ min 1.5, 1 + ξ2 + 3ξ3
1 ≤ y2 (ξ) ≤ 2 + ξ2 + 3ξ3 − y1 (ξ)
➥ ξ1 appears only on the left-hand side of the first constraint, so we can replace it
with its maximum value 0.5;
similarly, ξ2 , ξ3 appear only on the right-hand side of the second constraint, so we can
replace them with their minimum value 0.
➥ The optimal recourse is y1 (ξ) = y2 (ξ) = 1, that is a constant recourse.

More generally, if the uncertainty is constraint-wise, then the optimal recourse is


constant.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 5 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

RO is optimal for constraint-wise uncertainty II

( n o n o
max 1, 0.5 + ξ1 ≤ y1 (ξ) ≤ min 1.5, 1 + ξ2 + 3ξ3
1 ≤ y2 (ξ) ≤ 2 + ξ2 + 3ξ3 − y1 (ξ)
➥ ξ1 appears only on the left-hand side of the first constraint, so we can replace it
with its maximum value 0.5;
similarly, ξ2 , ξ3 appear only on the right-hand side of the second constraint, so we can
replace them with their minimum value 0.
➥ The optimal recourse is y1 (ξ) = y2 (ξ) = 1, that is a constant recourse.

More generally, if the uncertainty is constraint-wise, then the optimal recourse is


constant.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 5 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

RO is optimal for constraint-wise uncertainty II

( n o n o
max 1, 0.5 + ξ1 ≤ y1 (ξ) ≤ min 1.5, 1 + ξ2 + 3ξ3
1 ≤ y2 (ξ) ≤ 2 + ξ2 + 3ξ3 − y1 (ξ)
➥ ξ1 appears only on the left-hand side of the first constraint, so we can replace it
with its maximum value 0.5;
similarly, ξ2 , ξ3 appear only on the right-hand side of the second constraint, so we can
replace them with their minimum value 0.
➥ The optimal recourse is y1 (ξ) = y2 (ξ) = 1, that is a constant recourse.

More generally, if the uncertainty is constraint-wise, then the optimal recourse is


constant.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 5 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

RO is optimal for constraint-wise uncertainty II

( n o n o
max 1, 0.5 + ξ1 ≤ y1 (ξ) ≤ min 1.5, 1 + ξ2 + 3ξ3
1 ≤ y2 (ξ) ≤ 2 + ξ2 + 3ξ3 − y1 (ξ)
➥ ξ1 appears only on the left-hand side of the first constraint, so we can replace it
with its maximum value 0.5;
similarly, ξ2 , ξ3 appear only on the right-hand side of the second constraint, so we can
replace them with their minimum value 0.
➥ The optimal recourse is y1 (ξ) = y2 (ξ) = 1, that is a constant recourse.

More generally, if the uncertainty is constraint-wise, then the optimal recourse is


constant.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 5 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Piecewise-linear decision rules are optimal

For linear two-stage robust optimization problems, we can find an optimal recourse
that is piecewise-linear in the uncertainty ξ.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 6 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Contents

1 Fourier-Motzkin Elimination

2 Bender’s decomposition for two-stage problems

3 Linear Decision Rules

4 Piecewise-constant Decision Rules

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 6 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Two-stage robust linear problem setting


We consider the following two-stage robust linear problem:

min c ⊤ x + max d ⊤ yξ
x,y (·) ξ∈R

s.t. Aξ x + Bξ yξ ≤ bξ ∀ξ ∈ R
x ∈ X , yξ ≥ 0 ∀ξ ∈ R

Which can be reformulated as:

min c ⊤ x + V (x) where V (x) = max Ṽ (x, ξ)


x∈X ξ∈R

With n o
Ṽ (x, ξ) := min d ⊤ yξ Aξ x + Bξ yξ ≤ bξ
yξ ≥0

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 7 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Two-stage robust linear problem setting


We consider the following two-stage robust linear problem:

min c ⊤ x + max d ⊤ yξ
x,y (·) ξ∈R

s.t. Aξ x + Bξ yξ ≤ bξ ∀ξ ∈ R
x ∈ X , yξ ≥ 0 ∀ξ ∈ R

Which can be reformulated as:

min c ⊤ x + V (x) where V (x) = max Ṽ (x, ξ)


x∈X ξ∈R

With n o
Ṽ (x, ξ) := min d ⊤ yξ Aξ x + Bξ yξ ≤ bξ
yξ ≥0

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 7 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Bender’s decomposition for two-stage problems

The function Ṽ (·, ξ) is polyhedral.


The function V (x) is convex in x.
If R is a polyhedron, function V (x) is polyhedral.
➥ We can use a Bender’s decomposition approach to solve the problem.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 8 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Contents

1 Fourier-Motzkin Elimination

2 Bender’s decomposition for two-stage problems

3 Linear Decision Rules

4 Piecewise-constant Decision Rules

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 8 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Upper bound through decision rules

min max J(x, ξ, y (ξ))


x∈Rp ,y (·):R7→Rq ξ∈R

s.t. g (x, ξ, y (ξ)) ≤ 0 ∀ξ ∈ R

An idea is to restrict the recourse decision to be a parametric function (a.k.a decision


rule) of the uncertainty ξ,
y (ξ) = πθ (ξ)
and see the parameter θ as a first-stage decision variable:

min max J(x, ξ, πθ (ξ))


x∈Rp ,θ ξ∈R

s.t. g (x, ξ, πθ (ξ)) ≤ 0 ∀ξ ∈ R


V. Leclère Robust Optimization: Adaptive Case December 15, 2023 9 / 14
Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Upper bound through decision rules

min max J(x, ξ, y (ξ))


x∈Rp ,y (·):R7→Rq ξ∈R

s.t. g (x, ξ, y (ξ)) ≤ 0 ∀ξ ∈ R

An idea is to restrict the recourse decision to be a parametric function (a.k.a decision


rule) of the uncertainty ξ,
y (ξ) = πθ (ξ)
and see the parameter θ as a first-stage decision variable:

min max J(x, ξ, πθ (ξ))


x∈Rp ,θ ξ∈R

s.t. g (x, ξ, πθ (ξ)) ≤ 0 ∀ξ ∈ R


V. Leclère Robust Optimization: Adaptive Case December 15, 2023 9 / 14
Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Two-stage robust linear problem setting

min c ⊤ x + max d ⊤ yξ
x,y (·) ξ∈R

s.t. Aξ x + Bξ yξ ≤ bξ ∀ξ ∈ R
x∈ Rp+ , yξ ∈ Rq+ ∀ξ ∈ R

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 10 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Linear Decision Rules

We consider a linear decision rule:

yξ = V ξ + u

Thus, the two-stage problem becomes:

min c ⊤ x + max d ⊤ (V ξ + u)
x,u,V ξ∈R

s.t. Aξ x + Bξ V ξ + Bξ u ≤ bξ ∀ξ ∈ R
x ≥ 0, V ξ + u ≥ 0 ∀ξ ∈ R

As this problem is now one-stage, we can reformulate each constraint independently,


using the methods seen previously.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 11 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Linear Decision Rules

We consider a linear decision rule:

yξ = V ξ + u

Thus, the two-stage problem becomes:

min c ⊤ x + max d ⊤ (V ξ + u)
x,u,V ξ∈R

s.t. Aξ x + Bξ V ξ + Bξ u ≤ bξ ∀ξ ∈ R
x ≥ 0, V ξ + u ≥ 0 ∀ξ ∈ R

As this problem is now one-stage, we can reformulate each constraint independently,


using the methods seen previously.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 11 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Linear Decision Rules

Using LDR, we go from a two-stage problem to a one-stage problem, with


increased an number of variables.
Any LDR, optimal or not, yields a valid upper bound on the optimal value of the
two-stage problem.
For two-stage linear problems, with right-hand side uncertainty only, where R is a
simplex, LDR are optimal.
We can use LDR for multistage problems, non-anticipativity constraints being
represented through constraints on the matrix V .
LDR are not suitable for integer recourse variables.

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 12 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Contents

1 Fourier-Motzkin Elimination

2 Bender’s decomposition for two-stage problems

3 Linear Decision Rules

4 Piecewise-constant Decision Rules

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 12 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

Partition methodology

In order to simplify the problem, we can partition the uncertainty set R into a finite
number of subsets R1 , . . . , RK , and associate a recourse to each yk .
Thus, the two-stage problem becomes:

min max max J(x, ξ, yk )


x∈Rp ,(yk )k∈[K ] k∈K ξ∈Rk

s.t. g (x, ξ, yk ≤ 0 ∀k ∈ [K ], ∀ξ ∈ Rk

V. Leclère Robust Optimization: Adaptive Case December 15, 2023 13 / 14


Fourier-Motzkin Elimination Bender’s decomposition for two-stage problems Linear Decision Rules Piecewise-constant Decision Rules

D. Bertsimas, D. Brown, C. Caramanis


Theory and applications of robust optimization
Siam Review, 2011.
D. Bertsimas and D. Den Hertog
Robust and adaptive optimization
Dynamic Ideas, 2022.
BL Gorissen, I. Yanikoglu and D. den Hertog
A practical guide to robust optimization
Omega, 2015.
D. Bertsimas and M. Sim
The price of robustness
Operations research, 2004.
A. Ben Tal, L El Ghaoui, A. Nemirovski
Robust optimization
Springer, 2009.
V. Leclère Robust Optimization: Adaptive Case December 15, 2023 14 / 14

You might also like