Robust combinatorial optimization with

polyhedral uncertainty

Michael Poss
based on joint work with Marin Bougeret, Jérémy Omer, Artur Pessoa

October 2, 2024

1 / 29
Table of Contents

General results on polyhedral robust combinatorial optimization

Extension III: Application to decision-dependent information discovery

A crash course in open science

2 / 29
Given:
▶ Y ⊆ {0, 1}n : feasibility set of an optimization problem
▶ U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}: uncertainty polytope
▶ c: nominal cost vector

Solve:
min max u ⊤ y (Min-Max)
y ∈Y u∈U

Example of U:
u2

(0, 0) u1

3 / 29
Given:
▶ Y ⊆ {0, 1}n : feasibility set of an optimization problem
▶ U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}: uncertainty polytope
▶ c: nominal cost vector

Solve:
min max (c + u ′ )⊤ y (Min-Max)
y ∈Y u∈U

Example of U:
u2 u2′

c (0, 0)
u1′

(0, 0) u1

3 / 29
Given:
▶ Y ⊆ {0, 1}n : feasibility set of an optimization problem
▶ U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}: uncertainty polytope
▶ c: nominal cost vector

Solve:
min max (c + u ′ )⊤ y (Min-Max)
y ∈Y u∈U

Example of U:
u2 u2′

c (0, 0)
u1′

(0, 0) u1

3 / 29
Given:
▶ Y ⊆ {0, 1}n : feasibility set of an optimization problem
▶ U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}: uncertainty polytope
▶ c: nominal cost vector

Solve:
min max (c ′ + u ′′ )⊤ y (Min-Max)
y ∈Y u∈U

Example of U:
u2 u2′ u2′′
u1′′
c (0, 0)
u1′ (0, 0)

(0, 0) u1

3 / 29
 A 3D example
u3′′
Translated
 description:
U ′′ = u ′′ ∈ R3 | u1′′ + u2′′ + u3′′ ≤ 2, 0 ≤ u ′′ ≤ 1

u2′′
u1′′

4 / 29
 A 3D example
u3′′
Translated
 description:
U ′′ = u ′′ ∈ R3 | u1′′ + u2′′ + u3′′ ≤ 2, 0 ≤ u ′′ ≤ 1

u2′′
u1′′

“Natural”
n description: 8 symmetric copies of the polytope o
P3
U = u ∈ R3 | ci − 1 ≤ ui ≤ ci + 1, ∀i ∈ [3], i=1 |ui − c i | ≤ 2

4 / 29
 A 3D example
u3′′
Translated
 description:
U ′′ = u ′′ ∈ R3 | u1′′ + u2′′ + u3′′ ≤ 2, 0 ≤ u ′′ ≤ 1

u2′′
u1′′

“Natural”
n description: 8 symmetric copies of the polytope o
P3
U = u ∈ R3 | ci − 1 ≤ ui ≤ ci + 1, ∀i ∈ [3], i=1 |ui − c i | ≤ 2
( 3 3
X X
= u ∈ R3 | ci − 1 ≤ ui ≤ ci + 1, ∀i ∈ [3], (ui − ci ) ≤ 2, (−ui + ci ) ≤ 2,
i=1 i=1
X
(ui − ci ) − uj + cj ≤ 2, ∀I ∪ j = [3],
i∈I
)
X
(−ui + ci ) + uj − cj ≤ 2, ∀I ∪ j = [3]
i∈I 4 / 29
Hardness I: the 2-scenarios case is hard
Theorem
The robust selection problem is NP-hard even when |U| = 2.

Proof.

5 / 29
Hardness I: the 2-scenarios case is hard
Theorem
The robust selection problem is NP-hard even when |U| = 2.

Proof.
P
1. SELECTION PROBLEM: min ui
S⊆N,|S|=p i∈S

5 / 29
Hardness I: the 2-scenarios case is hard
Theorem
The robust selection problem is NP-hard even when |U| = 2.

Proof.
P
1. SELECTION PROBLEM: min ui
S⊆N,|S|=p i∈S
P
2. ROBUST SEL. PROB.: min max ui
S⊆N,|S|=p u∈U i∈S

5 / 29
Hardness I: the 2-scenarios case is hard
Theorem
The robust selection problem is NP-hard even when |U| = 2.

Proof.
P
1. SELECTION PROBLEM: min ui
S⊆N,|S|=p i∈S
P
2. ROBUST SEL. PROB.: min max ui
S⊆N,|S|=p u∈U i∈S
P P

3. PARTITION PROBLEM: min max ai , ai
S⊆N,|S|=|N|/2 i∈S i∈N\S

5 / 29
Hardness I: the 2-scenarios case is hard
Theorem
The robust selection problem is NP-hard even when |U| = 2.

Proof.
P
1. SELECTION PROBLEM: min ui
S⊆N,|S|=p i∈S
P
2. ROBUST SEL. PROB.: min max ui
S⊆N,|S|=p u∈U i∈S
P P

3. PARTITION PROBLEM: min max ai , ai
S⊆N,|S|=|N|/2 i∈S i∈N\S
|N| 1 2
4. Reduction: p = 2 , and U = {u , u } such that

2 X
ui1 = ai
and ui2 = ak − ai
P P P
|N| k
⇒ max ui = max ai , ai
u∈U i∈S i∈S i∈N\S

5 / 29
Hardness II: polytope is more general than 2 scenarios

Corollary
The robust selection problem is NP-hard when the number of rows of
A is part of the input.

Proof.

6 / 29
Hardness II: polytope is more general than 2 scenarios

Corollary
The robust selection problem is NP-hard when the number of rows of
A is part of the input.

Proof.
1. y ∗ ∈ arg min max (c + u)⊤ y ⇔ y ∗ ∈ arg min max (c + u)⊤ y
y ∈Y u∈U y ∈Y u∈ext(U)

6 / 29
Hardness II: polytope is more general than 2 scenarios

Corollary
The robust selection problem is NP-hard when the number of rows of
A is part of the input.

Proof.
1. y ∗ ∈ arg min max (c + u)⊤ y ⇔ y ∗ ∈ arg min max (c + u)⊤ y
y ∈Y u∈U y ∈Y u∈ext(U)
1 2 1 2
2. {u , u } is hard =⇒ conv({u , u }) is hard.

6 / 29
Hardness II: polytope is more general than 2 scenarios

Corollary
The robust selection problem is NP-hard when the number of rows of
A is part of the input.

Proof.
1. y ∗ ∈ arg min max (c + u)⊤ y ⇔ y ∗ ∈ arg min max (c + u)⊤ y
y ∈Y u∈U y ∈Y u∈ext(U)
1 2 1 2
2. {u , u } is hard =⇒ conv({u , u }) is hard.
3. conv({u 1 , u 2 }) can be described by 2n + 2 inequalities

6 / 29
Constant number of rows
U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}, s denotes the number of rows of A

Theorem
Solving (Min-Max) amounts to solve O(ns ) problems of the form
min c̄ ⊤ y
y ∈Y

Proof.

7 / 29
Constant number of rows
U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}, s denotes the number of rows of A

Theorem
Solving (Min-Max) amounts to solve O(ns ) problems of the form
min c̄ ⊤ y
y ∈Y

Proof.
1. dualize the maximization over u =⇒ dual variables (α, π) ∈ D(y )

7 / 29
Constant number of rows
U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}, s denotes the number of rows of A

Theorem
Solving (Min-Max) amounts to solve O(ns ) problems of the form
min c̄ ⊤ y
y ∈Y

Proof.
1. dualize the maximization over u =⇒ dual variables (α, π) ∈ D(y )
2. we can construct a set |D|∗ ∈ O(ns ) such that:
[ h i
Projα ext(D(y )) ⊆ D ∗
x∈{0,1}n

=⇒ enumerate α ∈ D ∗

7 / 29
Constant number of rows
U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}, s denotes the number of rows of A

Theorem
Solving (Min-Max) amounts to solve O(ns ) problems of the form
min c̄ ⊤ y
y ∈Y

Proof.
1. dualize the maximization over u =⇒ dual variables (α, π) ∈ D(y )
2. we can construct a set |D|∗ ∈ O(ns ) such that:
[ h i
Projα ext(D(y )) ⊆ D ∗
x∈{0,1}n

=⇒ enumerate α ∈ D ∗
3. variables πk can be substituted by max(yk − f (α), 0)

7 / 29
Constant number of rows
U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d}, s denotes the number of rows of A

Theorem
Solving (Min-Max) amounts to solve O(ns ) problems of the form
min c̄ ⊤ y
y ∈Y

Proof.
1. dualize the maximization over u =⇒ dual variables (α, π) ∈ D(y )
2. we can construct a set |D|∗ ∈ O(ns ) such that:
[ h i
Projα ext(D(y )) ⊆ D ∗
x∈{0,1}n

=⇒ enumerate α ∈ D ∗
3. variables πk can be substituted by max(yk − f (α), 0)
4. max(yk − f (α)) = yk max(1 − f (α)) + (1 − yk ) max(−f (α), 0)

7 / 29
Extension I: non-linear objective
The previous results extend to:

min max(c + u)⊤ f (x)

y ∈Y u∈U

where fk (x) ∈ {0, Fk } for each y ∈ Y.

Examples
▶ max-cut with uncertain weights:
X
max min wij yi (1 − yj ),
y ∈Y w ∈U
{i,j}∈E

with Fij = 1.

8 / 29
Extension I: non-linear objective
The previous results extend to:

min max(c + u)⊤ f (x)

y ∈Y u∈U

where fk (x) ∈ {0, Fk } for each y ∈ Y.

Examples
▶ max-cut with uncertain weights:
X
max min wij yi (1 − yj ),
y ∈Y w ∈U
{i,j}∈E

with Fij = 1.
▶ minimizing weight of falling jobs
X
min max wi Ui (y ),
y ∈Y w ∈U
i∈J

with Ui (y ) = 1 iff job i fails in schedule y .

8 / 29
Extension II: constraint uncertainty
Define:
min c ⊤y
Ye = Y ∩ {(a + u)⊤ y ≤ h, ∀u ∈ U}
s.t. (a + u)⊤ y ≤ h, ∀u ∈ U
y ∈Y (1) Y(α)
e = Y ∩ {a(α)⊤ y ≤ h}

Theorem
[
Ye = Y(α)
e
α∈D ∗

9 / 29
Extension II: constraint uncertainty
Define:
min c ⊤y
Ye = Y ∩ {(a + u)⊤ y ≤ h, ∀u ∈ U}
s.t. (a + u)⊤ y ≤ h, ∀u ∈ U
y ∈Y (1) Y(α)
e = Y ∩ {a(α)⊤ y ≤ h}

Theorem
[
Ye = Y(α)
e
α∈D ∗

Corollary
Solving (1) amounts to solve O(ns ) problems of the form

min c ⊤y
s.t. ā⊤ y ≤ h, ∀u ∈ U
y ∈Y

9 / 29
Extension II: constraint uncertainty - D.-W. reformulation
master pricing
z}|{ z}|{
Y = YM ∩ YP

10 / 29
Extension II: constraint uncertainty - D.-W. reformulation
master pricing
z}|{ z}|{
Y = YM ∩ YP

YeP = Y P ∩{(a+u)⊤ y ≤ h, ∀u ∈ U}
Master:
!
X
⊤
min c λs xs
s
X
s.t. λs xs ∈ Y M
s
λ∈Λ
n o
Pricing: min c ⊤ y | x ∈ YeP

10 / 29
Extension II: constraint uncertainty - D.-W. reformulation
master pricing
z}|{ z}|{
Y = YM ∩ YP

YeP (α) = Y P ∩ {a(α)⊤ y ≤ h}

YeP = Y P ∩{(a+u)⊤ y ≤ h, ∀u ∈ U}
Master:
Master: !
! X X
⊤
X min c λs (α)xs
min c⊤ λs xs α∈D ∗ s
s X X
X
M s.t. λs (α)xs ∈ Y M
s.t. λs xs ∈ Y
α∈D ∗ s
s X
λ∈Λ λ(α) ∈ Λ
α∈D ∗
n o
Pricing: min c ⊤ y | x ∈ YeP n o
Pricing: min c ⊤ y | x ∈ YeP (α)

See Pessoa et al. [2021] for pre-processing techniques reducing the

number of α ∈ D ∗
10 / 29
Table of Contents

General results on polyhedral robust combinatorial optimization

Extension III: Application to decision-dependent information discovery

A crash course in open science

11 / 29
Step by step definition of the problem
Note: for the sake of simplicity, we consider
( )
n
P
∆= δ∈R | δi ≤ 1, 0 ≤ δ ≤ 1
i∈[n]

See Omer et al. [2024] for the generalization to U.

Decision dependent information discovery (DDID)
Before choosing y, one may ask for q deviations to be revealed:
▶ worst-case realization of those q deviations is considered,
▶ y is chosen after those q observations.
⇒ddid seeks for the optimal choice of the q observations.

12 / 29
Step by step definition of the problem
Note: for the sake of simplicity, we consider
( )
n
P
∆= δ∈R | δi ≤ 1, 0 ≤ δ ≤ 1
i∈[n]

See Omer et al. [2024] for the generalization to U.

Decision dependent information discovery (DDID)
Before choosing y, one may ask for q deviations to be revealed:
▶ worst-case realization of those q deviations is considered,
▶ y is chosen after those q observations.
⇒ddid seeks for the optimal choice of the q observations.

X
z DDID = min max min max (ci + di δi )yi , (ddid)
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]

▶ W = {w ∈ {0, 1}n |
P
i wi = q},

▶ ∆(w , δ̄) = δ ∈ ∆ | w ◦ δ = w ◦ δ̄ : deviations cannot change once
observed.
12 / 29
Applications
Any planning process were one may investigate the parameters before
taking the actual decision:

organ transplant (e.g. kidney exchange) underground works

Y = {short cycles, short paths} Y = {trees, . . .}
Carvalho et al. [2021] Focke et al. [2020]

Many more applications in Vayanos et al. [2020]

13 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ How to solve the following ?

X
max (ci + di δi )y i
δ∈∆(w ,δ̄)
i∈[n]

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ How to solve the following ?

X
max (ci + di δi )y i
δ∈∆(w ,δ̄)
i∈[n]

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ How to solve the following ?

X
max (ci + di δi )y i
δ∈∆(w ,δ̄)
i∈[n]

▶ And the following ?

X
min max (ci + di δi )y i .
y ∈Y δ∈∆(w ,δ̄)
i∈[n]

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ How to solve the following ?

X
max (ci + di δi )y i
δ∈∆(w ,δ̄)
i∈[n]

▶ And the following ?

X
min max (ci + di δi )y i .
y ∈Y δ∈∆(w ,δ̄)
i∈[n]

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ How to solve the following ?

X
max (ci + di δi )y i
δ∈∆(w ,δ̄)
i∈[n]

▶ And the following ?

X
min max (ci + di δi )y i .
y ∈Y δ∈∆(w ,δ̄)
i∈[n]

Theorem ([Bertsimas and Sim, 2003])

X X
min max (ci + di δi )y i = min Γd ℓ + min ciℓ y i ,
y ∈Y δ∈∆ ℓ∈[n] y ∈Y
i∈[n] i∈[n]

ℓ ℓ
where d and c , ℓ ∈ [n], follow simple formulas.
14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ So the following is easy:

X
min max (ci + di δi )y i .
y ∈Y δ∈∆(w ,δ̄)
i∈[n]

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ So the following is easy:

X
min max (ci + di δi )y i .
y ∈Y δ∈∆(w ,δ̄)
i∈[n]

▶ What about the following ?

X
max min max (ci + di δi )y i .
δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]

14 / 29
Robust combinatorial optimization
Let’s decompose X
min max min max (ci + di δi )y i .
w ∈W δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
n o
Remember: ∆ = δ ∈ [0, 1]n | i∈[n] δi ≤ Γ .
P

▶ So the following is easy:

X
min max (ci + di δi )y i .
y ∈Y δ∈∆(w ,δ̄)
i∈[n]

▶ What about the following ?

X
max min max (ci + di δi )y i .
δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]

⇒ This is our main contribution. We study

X
Φ(w ) = max min max (ci + di δi )y i ,
δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]
under two assumptions:
Y ⊆ {0, 1}n and conv(Y) = P

14 / 29
Linear formulation of the adversary problem I
For a given choice of observations w ∈ W, the (outer) adversary problem
chooses the revealed deviations δ̄:
X
Φ(w ) = max min max (ci + di δi )y i ,
δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]

15 / 29
Linear formulation of the adversary problem I
For a given choice of observations w ∈ W, the (outer) adversary problem
chooses the revealed deviations δ̄:
X
Φ(w ) = max min max (ci + di δi )y i ,
δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]

¯
Let us “simplify” ∆(w , δ̄) to ∆:
X
Φ(w ) = max min max (c̄i + d¯i δi )y i ,
¯
δ̄∈∆ y ∈Y δ∈∆
i∈[n]

where:
▶ c̄i and Γ̄ are affine functions of δ̄
¯ = δ ∈ [0, 1]n | P δi ≤ Γ̄

▶ ∆ i
▶ d¯ is independent of δ̄

15 / 29
Linear formulation of the adversary problem I
For a given choice of observations w ∈ W, the (outer) adversary problem
chooses the revealed deviations δ̄:
X
Φ(w ) = max min max (ci + di δi )y i ,
δ̄∈∆ y ∈Y δ∈∆(w ,δ̄)
i∈[n]

¯
Let us “simplify” ∆(w , δ̄) to ∆:
X
Φ(w ) = max min max (c̄i + d¯i δi )y i ,
¯
δ̄∈∆ y ∈Y δ∈∆
i∈[n]

where:
▶ c̄i and Γ̄ are affine functions of δ̄
¯ = δ ∈ [0, 1]n | P δi ≤ Γ̄

▶ ∆ i
▶ d¯ is independent of δ̄
1. Epigraphic formulation:
Φ(w ) = max ω
δ̄∈∆
X
s.t. ω ≤ min max (c̄i + d¯i δi )y i
¯
y ∈Y δ∈∆
i∈[n]

15 / 29
Linear formulation of the adversary problem II
1. Epigraphic formulation:

Φ(w ) = max ω
δ̄∈∆
X
s.t. ω ≤ min max (c̄i + d¯i δi )y i
¯
y ∈Y δ∈∆
i∈[n]

2. B&S theorem: rhs is equivalent to solving n + 1 independent

problems.

max ω
δ̄∈∆
X
s.t. ω ≤ min Γ̄d¯ℓ + min c̄iℓ y i ,
ℓ∈[n]0 y ∈Y
i

16 / 29
Linear formulation of the adversary problem II
1. Epigraphic formulation:

Φ(w ) = max ω
δ̄∈∆
X
s.t. ω ≤ min max (c̄i + d¯i δi )y i
¯
y ∈Y δ∈∆
i∈[n]

2. B&S theorem: rhs is equivalent to solving n + 1 independent

problems.

max ω
δ̄∈∆
X
s.t. ω ≤ min Γ̄d¯ℓ + min c̄iℓ y i ,
ℓ∈[n]0 y ∈Y
i

max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + min c̄iℓ y i , ℓ ∈ [n]0
y ∈Y
i

16 / 29
Linear formulation of the adversary problem III
2. B&S theorem: rhs is equivalent to solving n + 1 independent problems.
max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + min c̄iℓ y i , ℓ ∈ [n]0
y ∈Y
i

17 / 29
Linear formulation of the adversary problem III
2. B&S theorem: rhs is equivalent to solving n + 1 independent problems.
max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + min c̄iℓ y i , ℓ ∈ [n]0
y ∈Y
i

3. Dualization: use that conv(Y) = P to dualize the minimization.

Φ(w ) = max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + maxb T λℓ − πℓ,i , ℓ ∈ [n]0
π,λ
i∈[n]

s.t.(B·,i )T λℓ − πℓ,i ≤ c̄iℓ , ∀ℓ ∈ [n]0 , ∀i

λℓ , πℓ ≥ 0, ∀ℓ ∈ [n]0

17 / 29
Linear formulation of the adversary problem III
2. B&S theorem: rhs is equivalent to solving n + 1 independent problems.
max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + min c̄iℓ y i , ℓ ∈ [n]0
y ∈Y
i

3. Dualization: use that conv(Y) = P to dualize the minimization.

Φ(w ) = max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + maxb T λℓ − πℓ,i , ℓ ∈ [n]0
π,λ
i∈[n]

s.t.(B·,i )T λℓ − πℓ,i ≤ c̄iℓ , ∀ℓ ∈ [n]0 , ∀i

λℓ , πℓ ≥ 0, ∀ℓ ∈ [n]0
⇔ Φ(w ) = max ω
δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + b T λℓ − πℓ,i , ℓ ∈ [n]0
i∈[n]

(B·,i )T λℓ − πℓ,i ≤ c̄iℓ , ∀ℓ ∈ [n]0 , ∀i

λℓ , πℓ ≥ 0, ∀ℓ ∈ [n]0

17 / 29
Linear formulation of the adversary problem IV

We have reformulated min Φ(w ) as:

w ∈W

min max ω
w ∈W δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + b T λℓ − πℓ,i , ℓ ∈ [n]0
i∈[n]

(B·,i )T λℓ − πℓ,i ≤ c̄iℓ , ∀ℓ ∈ [n]0 , ∀i

λℓ , πℓ ≥ 0, ∀ℓ ∈ [n]0

18 / 29
Linear formulation of the adversary problem IV

We have reformulated min Φ(w ) as:

w ∈W

min max ω
w ∈W δ̄∈∆
X
s.t. ω ≤ Γ̄d¯ℓ + b T λℓ − πℓ,i , ℓ ∈ [n]0
i∈[n]

(B·,i )T λℓ − πℓ,i ≤ c̄iℓ , ∀ℓ ∈ [n]0 , ∀i

λℓ , πℓ ≥ 0, ∀ℓ ∈ [n]0

The formulation of the ddid is finally obtained by

▶ dualization of inner maximization problem,
▶ linearization of all the hidden products with the w .

18 / 29
Main result: Compact MILP formulation for ddid

 
X X X X
min r αℓ u ℓ + ci y ℓ,i + βℓ,i y 0ℓ,i  + di σ i
ℓ∈[n]0 i∈[n] i∈[n] i∈[n]
X
s.t. uℓ = 1
ℓ∈[n]0
X X
ai σi ≥ −ai αℓ u ℓ + y ℓ,i − (1 − w i ), ∀i ∈ [n]
ℓ∈[n]0 ℓ∈[n]0

By ℓ ≥ u ℓ b, ∀ℓ ∈ [n]0
y ℓ,i ≤ u ℓ , ∀ℓ ∈ [n]0 , i ∈ [n]
y 0ℓ,i ≥ y ℓ,i − w i , ∀ℓ ∈ [n]0 , i ∈ [n]
0
w ∈ W, u, y , y , σ ≥ 0.

19 / 29
Selection problem
Y = {select p = n/10 items among n to minimize total cost}
Vayanos et al. [2020]: K -adaptability heuristic: decision-maker must
chooses K strategies y 1 , . . . , y K among which one is chosen after
revealing δ̄.
▶ our own implementation
▶ budget uncertainty (L = 1)
▶ K -adaptability reformulations grow linearly in L and K

n T gap
n K T gap
10 6 0.14
2 7 10
10 20 10 0.26
3 24 11
30 26 0.39
2 46 7
15 40 49 0.16
3 3275 9
50 160 0.22
(a) K -adaptability MILP reformulation.
(b) Our exact MILP reformulation.

Note: times are in centiseconds

20 / 29
Orienteering problem

Y = {elementary paths with maximum time constraint}

▶ CB : exact algorithm: column and constraint generation to solve the
adversary problem embedded in combinatorial Benders’ cut
algorithm [Paradiso et al., 2022].
▶ BP : our branch-and-price algorithm

instance Opt Solved at root

CB BP BP
TS2N10 100% 100% 89%
TS1N15 100% 100% 90%
TS3N16 100% 100% 83%
TS2N19 76% 100% 73%
TS1N30 41% 96% 61%
TS3N31 37% 90% 80%

21 / 29
Conclusion

U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d} , L is the number of rows of A

Take-away messages
▶ min-max: not harder than nominal problem if L is constant
▶ ddid: If conv(Y) = P is a compact polyhedron and L is constant
▶ Computing Φ is a compact LP
▶ Solving ddid is a compact MILP

22 / 29
Conclusion

U = {u ∈ Rn | Au ≤ b, 0 ≤ u ≤ d} , L is the number of rows of A

Take-away messages
▶ min-max: not harder than nominal problem if L is constant
▶ ddid: If conv(Y) = P is a compact polyhedron and L is constant
▶ Computing Φ is a compact LP
▶ Solving ddid is a compact MILP

Perspectives
▶ Complexity: it is not known yet whether ddid can be NP-Hard
when the nominal problem is polynomial ⇒ thesis of Xiaoyu Chen
▶ Numerically: decomposition is promising when conv(Y) ̸= P

22 / 29
Table of Contents

General results on polyhedral robust combinatorial optimization

Extension III: Application to decision-dependent information discovery

A crash course in open science

23 / 29
What is open science?
Assuming our research is useful:
▶ Make the results of science available to everybody
▶ Help disseminating science to society

Open data ©
Benchmarks (MIPlib, QPlib, netlib. . .), real data from industrial
applications

Open code §
§§ Most published algorithms are not reproducible!
© Some journals enforce reproducibility (IJOC, MPC, OJMO, OR)

Open publications §
© Optimization online, arxiv
§ Many papers not available online
§§ Ever-increasing publication fees (cost ∼ 100M¤ annually in France)
24 / 29
The beginnings ...
Back in the days, publishing was expensive!

25 / 29
Internet and LATEX
▶ The advent of electronic publishing and dissemination via the Web
and the use of LATEXreduced significantly the operating costs of
publishers
▶ Led to nearly open and free publications?
▶ In fact no:

Note: Publishers ask at least 2000¤ for Open Access. Its real cost varies
between 3¤and 800¤ (when heavy typesetting is needed).
26 / 29
The uprising

Taken from the presentation of Marie Farge

27 / 29
Examples of fair journals/conferences
Machine learning, Artificial intelligence
▶ JAIR
▶ Journal of Machine Learning Research (JMLR)
▶ Transactions on Machine Learning Research (TMLR)
Graph theory, algorithmics
▶ Advances in Combinatorics
▶ TheoretiCS
▶ Theory of Computing
▶ Innovations in Graph Theory
And many conferences: LIPICs (mainly theoretical CS), and more . . .

28 / 29
Examples of fair journals/conferences
Machine learning, Artificial intelligence
▶ JAIR
▶ Journal of Machine Learning Research (JMLR)
▶ Transactions on Machine Learning Research (TMLR)
Graph theory, algorithmics
▶ Advances in Combinatorics
▶ TheoretiCS
▶ Theory of Computing
▶ Innovations in Graph Theory
And many conferences: LIPICs (mainly theoretical CS), and more . . .

Mathematical optimization
▶ Open Journal of Mathematical Optimization (OJMO)
▶ Area editors: G. Bayraksan, R. Luke, J. Malick, S. Pokutta
▶ indexed in most databases, Q2 at scimago
▶ fast track for short papers
▶ enforces reproducibility!
28 / 29
References I
Dimitris Bertsimas and Melvyn Sim. Robust discrete optimization and network flows.
Math. Program., 98(1-3):49–71, 2003. doi: 10.1007/s10107-003-0396-4. URL
https://doi.org/10.1007/s10107-003-0396-4.
Margarida Carvalho, Xenia Klimentova, Kristiaan Glorie, Ana Viana, and Miguel
Constantino. Robust models for the kidney exchange problem. INFORMS J.
Comput., 33(3):861–881, 2021. doi: 10.1287/IJOC.2020.0986. URL
https://doi.org/10.1287/ijoc.2020.0986.
Jacob Focke, Nicole Megow, and Julie Meißner. Minimum spanning tree under
explorable uncertainty in theory and experiments. ACM J. Exp. Algorithmics, 25:
1–20, 2020. doi: 10.1145/3422371. URL https://doi.org/10.1145/3422371.
Jérémy Omer, Michael Poss, and Maxime Rougier. Combinatorial Robust Optimization
with Decision-Dependent Information Discovery and Polyhedral Uncertainty. Open
Journal of Mathematical Optimization, 5:5, 2024. doi: 10.5802/ojmo.33. URL
https://ojmo.centre-mersenne.org/articles/10.5802/ojmo.33/.
Rosario Paradiso, Angelos Georghiou, Said Dabia, and Denise Tönissen. Exact and
approximate schemes for robust optimization problems with decision dependent
information discovery. arXiv preprint arXiv:2208.04115, 2022.
Artur Alves Pessoa, Michael Poss, Ruslan Sadykov, and François Vanderbeck.
Branch-cut-and-price for the robust capacitated vehicle routing problem with
knapsack uncertainty. Oper. Res., 69(3):739–754, 2021. doi:
10.1287/opre.2020.2035. URL https://doi.org/10.1287/opre.2020.2035.
Phebe Vayanos, Angelos Georghiou, and Han Yu. Robust optimization with
decision-dependent information discovery. arXiv preprint arXiv:2004.08490, 2020.
29 / 29