NetOpt2020 2020 01 20 Arie

Robust Network Optimization
Arie M.C.A. Koster

koster@math2.rwth-aachen.de
9th Winter School on Network Optimization – 20 January 2020

Today’s Program
Introduction to Robust Optimization (1st lecture)

Network Design under Demand Uncertainty (1st & 2nd lecture)
Extensions of Robust Optimization (2nd lecture)
Exercises/Research Questions (Afternoon workshop)
Arie Koster – RWTH Aachen University 2 / 101

Part I
Robust Optimization

Outline
1 Motivation: 2 Examples
2 Chance-Constrained Programming
3 Robust Optimization
4 Γ-Robust Optimization

Ex. 1: Telecom. Network Design
Network Design: Capacity Installation, Routing, Survivability

Based on demand predictions ... design as good as the prediction
How robust is a network design?
Demand uncertainties Traffic fluctuations in the US abilene Internet2

network in time intervals of 5 minutes during one week:
Traffic fluctuates heavily between node-pairs
Load of links will fluctuate alike
To avoid congestion, demand is overestimated by, e.g., ≥ 300%
Can we do better?
Alternative Approaches
Lower overestimation
The network is designed such that capacities are as small as possible; traffic
fluctuations might result in high network congestion
Stochastic Programming
Network design has to be computed for many scenarios; high computational
effort
Multi-period Network Design

Many traffic matrices have to be considered simultaneously; high
computational effort

Ex. 2: Combined-Heat-and-Power Plants
Simultaneous production of heat and power in exchange for fuel
Source: ProCom
Heat can be stored for future use, power cannot be stored
Power has to be bought/sold at

day-ahead market,
but demand for heat (and power) is
uncertain!

Outline

Chance-Constrained Programming
Chance-constrained Optimization Problem (COP)

Find among all solutions that satisfy all constraints with high probability a
solution with optimal objective value.
How to solve a COP?

Stochastic Optimization
I Modelling with random variables
I Quite challenging to solve resulting problems
I Probability distribution have to be determined
Robust Optimization
I Uncertainty comes from a known set, the uncertainty set
I No information on probability distribution needed
I Seeks for solution with best worst-case objective guarantee

Chance-Constrained Programming
Chance-Constrained Linear Programming with joint constraintswith

individual constraints
min cT x
s.t. Ax ≤ bP (Ax ≤ b) ≥ 1 − P (Ai x ≤ bi ) ≥ 1 − i ∀i = 1, . . . , m
x ≥0
with Entries of A, b and/or c are not constant but random variables

Example
Chance-Constrained Knapsack:
Knapsack with n Items, profits ci , uncertain weights ai , and capacity b
n
X
max ci xi
i=1
n
!
X
s.t. P ai x i ≤ b ≥1−
i=1
x ∈ {0, 1}n
How to solve this problem?

Assumption: Weights are independently and normally distributed with
expectation mi and standard deviation σi .

Example (cont.)
Assumption: Weights are independently and normally distributed with
expectation mi and standard deviation σi .
 
n
! Pn Pn
(a x − mi xi ) b− mi xi 
qP i i
X
P ai xi ≤ b = P  i=1 ≤ qP i=1
n 2 2 n 2 2
i=1 i=1 σi xi i=1 σi xi
 
b − ni=1 mi xi 
P
= P Z ≤ qP
 ≥1−
n 2x 2
σ
i=1 i i
Pn
i=1 (ai xi −mi xi )
√
with Z = Pn 2 2
i=1 σi xi
Let Φ(.) be the cumulative distribution function of the standard normal
distribution. Then,
b − ni=1 mi xi
P
qP ≥ Φ−1 (1 − )
n 2 2
i=1 σi xi

Example (cont.)
b − ni=1 mi xi
P
qP ≥ Φ−1 (1 − )
n 2 2
i=1 σi xi
If 1 − > 0.5, Φ−1 (1 − ) > 0 and the chance constrained knapsack can be
reformulated as
n
X
min ci xi
i=1
v
u n n
−1
uX X
s.t. Φ (1 − )t σi2 xi2 + mi xi ≤ b
i=1 i=1
n
x ∈ {0, 1}
After relaxing the integrality of x, a second order cone problem remains,

which can be solved in polynomial time.
Outline

Uncertain LPs
Observation
In the example, normal distribution of the weights was assumed. What if,
the weights are distributed differently, or unknown?
Uncertain Linear Program

An Uncertain Linear Optimization problem (ULO) is a collection of linear
optimization problems (instances)
n o
min{c T x : Ax ≤ b}
(c,A,b)∈U
where all input data stems from an uncertainty set U ⊂ Rn × Rm×n × Rm .

Robust Counterpart
n o
ULO min{c T x : Ax ≤ b}
(c,A,b)∈U
Robust feasible solution

A vector x ∈ Rn is robust feasible for ULO if
Ax ≤ b ∀(c, A, b) ∈ U
Robust solution value

Given a vector x ∈ Rn , the robust solution value ĉ(x) is defined as
ĉ(x) := sup cT x
(c,A,b)∈U
Robust Counterpart
The robust counterpart of an ULO is the optimization problem
min {ĉ(x) : x is robust feasible}

Example
n o
Let min{c T x : Ax ≤ b, x ≥ 0} be an ULO with uncertain
(c,A,b)∈U
right-hand-side
b ∈ [b̄, b̄ + b̂]
uncertain matrix A,
aij ∈ [āij , āij + âij ]
but certain objective vector c.
The robust counterpart can be written as
min{c T x : (Ā + Â)x ≤ b̄, x ≥ 0}

Robust Counterpart
Observation
If the objective is certain, the robust counterpart can be constructed
row-wise, i.e.,
keep the objective
replace every constraint aiT x ≤ bi by its robust counterpart
aiT x ≤ bi ∀(ai , bi ) ∈ Ui
where
n o
Ui := (ãi , b̃i ) ∈ Rn+1 : ∃(A, b) ∈ U with Ai. = ãi , bi = b̃i
Note: the robust counterpart does not change if Û = U1 × U2 × . . . × Um

instead of U is used.
Wlog: Objective vector c is certain!

Uncertainty Sets
How to define the uncertainty sets?

Uncertainty set is an ellipsoid [10], e.g.,
Ui = (a, b) ∈ Rn+1 : k(a, b) − (ā, b̄)k < κ

Uncertainty set is an polyhedron, e.g.,
Ui = (a, b) ∈ Rn+1 : D · (a, b) ≤ d

with D ∈ Rk×n , d ∈ Rk for some k ∈ N [2].

equivalent: set of discrete scenarios (extreme points of polyhedron)
special case: Γ-Robustness

Outline

Robust Optimization Approach
Simplifying assumption: b is certain
Uncertainty Set by Bertsimas & Sim [11, 12]: Let āij ∈ R, âij ≥ 0 be given,
and Γ ∈ R+ a parameter.
Ui (Γ) = {ai ∈ Rn : aij = āij + âij zij ∀j = 1, . . . , n, zi ∈ Zi (Γ)}
with
 
 n
X 
Zi (Γ) = zi ∈ Rn : |zij | ≤ 1 ∀j = 1, . . . , n, |zij | ≤ Γ
 
j=1
Stated otherwise:
nominal values āij and deviations âij
aij ∈ [āij − âij , āij + âij ]
Sum of relative deviations from the nominal values is bounded by Γ
“At most Γ many entries might deviate from their nominal value”
Example
   
2 1
ā = 5 , â = 2, Γ = 2
  
5 1
5
a3
4 6
1
1.5 2 4 a2
2.5 3
a1 Provided by Manuel Kutschka

Γ-Robust Counterpart
Robust Counterpart
X n
min ci xi
i=1
 
n
X Xn
s.t. āij xj + max  âij zij xj  ≤ bi i = 1, . . . , m
zi ∈Zi (Γ)
j=1 j=1
x ≥0
Observation
Since Zi defines a (bounded) polyhedron, only the extreme points have to
be treated.
For Γ ∈ Z+ :
 
n
X X
āij xj + max  âij xj  ≤ bi (1)
S⊆{1,...,n}:|S|≤Γ
j=1 j∈S

Γ-Robustness – Theory
Theorem 1 (Bertsimas & Sim [12])

Let x ? be an optimal solution of the Γ-robust counterpart. If aij ,
j = 1, . . . , n, are independent and symmetric distributed random variables in
[āij − âij , āij + âij ], then
n
!
X
P ai xi? > b ≤ B(n, Γ)
i=1
with
Γ
lim B(n, Γ) = 1 − Φ √
n→∞ n
where Φ(.) is the CDF of the standard normal distribution.

Γ−1
Instead of the limit: B(n, Γ) ≈ 1 − Φ √
n

Choice of Γ
Choice of Γ as a function of n so that the

probability of constraint violation is less than p%:
Γ
n p=1 p = 0.5 p = 0.1
5 5.0 5.0 5.0
10 8.4 9.1 10.0
100 24.3 26.8 31.9
200 33.9 37.4 44.7
1,000 74.6 82.5 98.7
2,000 105.0 116.2 139.2
Note: Result is independent of actual distribution of random variables aij ,

only symmetry and independence are required.

Γ-Robustness – Practice
Questions to be answered:
1. How to solve the robust counterpart?
⇒ General answer first, more detailed for Robust Network Design
2. Optimal value as function of Γ?
⇒ To be discussed for Robust Network Design
3. How to determine the nominal values āij and deviations âij ?
⇒ To be discussed for Robust Network Design
4. How robust is the obtained solution in practice?
⇒ To be discussed for Robust Netowrk Design

Solving the robust counterpart
 
n
X X
āij xj + max  âij xj  ≤ bi
S⊆{1,...,n}:|S|≤Γ
j=1 j∈S
Observations:
Inequality (1) can be linearized by
X X
āij xj + (āij + âij )xj ≤ bi ∀S ⊆ {1, . . . , n}, |S| ≤ Γ (2)
j6∈S j∈S
This number of inequalities is exponential if Γ = O(n)

Separation can be done in polynomial time
Alternatively, a compact formulation can be obtained via dualization

Separation of robust inequalities
Given x ? , find a subset S ⊆ {1, . . . , n} with |S| ≤ Γ such that
X X n
X X n
X
āij xj? + ?
(āij + âij )xj ?
āij xj + ?
âij xj > bi − āij xj?
j6∈S j∈S j=1 j∈S j=1
Separation problem:
n
X
ZSEP = max âij xj? zj
j=1
Xn
s.t. zj ≤ Γ
j=1
zj ∈ {0, 1}0 ≤ zj ≤ 1
If ZSEP > bi − nj=1 āij xj? , add robust inequality (2) for S = {j : zj = 1}.
P
Optimization = Separation implies polynomial solvability of LP
Compact Reformulation
 
X
Let βi (x, Γ) = max  âij xj , and hence (1) reads
S⊆{1,...,n}:|S|≤Γ
j∈S
n
X
āij xj + βi (x, Γ) ≤ bi
j=1
Given x ? , βi (x ? , Γ) is the optimization problem

n
X n
X
βi (x ? , Γ) = max âij xj? zj = minΓπi + ρij
j=1 j=1
n
X
s.t. zj ≤ Γ s.t.πi + ρij ≥ âij xj? ∀j = 1, . . . , n
j=1
0 ≤ zj ≤ 1 πi , ρij ≥ 0

Thus, (1) now reads

 
Xn  n
X 
āij xj + min Γπi + ρij : πi + ρij ≥ âij xj ∀j, πi ≥ 0, ρij ≥ 0 ≤ bi
 
j=1 j=1
or equivalently
n
X n
X
āij xj + Γπi + ρij ≤ bi
j=1 j=1
πi + ρij ≥ âij xj ∀j = 1, . . . , n
πi ≥ 0, ρij ≥ 0

Part II
Network Design under Demand

Uncertainty

Outline
5 Robust Network Design

6 Robust Cutset Inequalities
7 Back to practice

Design of Telecommunication Networks
given a potential network topology,
demand forecast,
link modules with
capacities/costs
find a hardware configuration,
and a routing,
such that demands are satisfied and total
installation cost is minimised.
discrete decisions:
Variations / Extensions:
Single path routing Node hardware (switching capacity)
Integer routing Wavelength assignment
Survivability requirements Multi-layer scenarios
Deterministic Network Design
yijk = fraction of demand k ∈ K routed along arc (i, j) ∈ A (yek := yijk + yjik ).
xe = number of link capacity modules to be installed on link e ∈ E .
Integer Linear Programming formulation:
X
min κe xe (3)
e∈E

X 1
 i = s(k)
k k
s.t. (yij − yji ) = −1 i = t(k) , ∀i ∈ V , k ∈ K (4)

{i,j}∈δ(i) 
0 else
X
d k yek ≤ Cxe , ∀e ∈ E (5)
k∈K
y, x ≥ 0 (6)
|E |
x ∈Z (7)

Network and traffic properties of data
Network Abilene Géant Germany17 Germany50

# nodes 12 22 17 50
# links 15 36 26 89
# demands 66 231 136 1044
available traffic period 6 months 4 months 1 day 1 day
traffic granularity 5 min 15 min 5 min 5 min
# available traffic matrices 48 095 10 737 288 288
# traffic matrices used 2 × 8 064 2 × 2 688 288 288
instances Abilene1 Geant1 Germany17 Germany50
Abilene2 Geant2
d¯k = geometric mean of 1 week / 1 day

dˆk = 95% percentile of 1 week / 1 day

Γ-Robust Network Design
Let d k be a random variable with nominal d¯k and deviation dˆk ;

d k ∈ [0, d¯k + dˆk ]
Goal: Network satisfying Γ deviated demands simultaneously
Only capacity constraints (5) change to:
 
X X 
d¯k yek + max dˆk yek ≤ Cxe , (5’)
Q⊆K ,|Q|≤Γ  
k∈K k∈Q
All other coefficients are deterministic.

Linearization by an exponential number of constraints.

Given y , subset Q ⊆ K maximizing the traffic can be computed by linear

programming:
 
X  X
max dˆk yek = max dˆk yek zek
Q⊆K ,|Q|≤Γ  
k∈Q k∈K
X
s.t. zek ≤ Γ
k∈K
0 ≤ zek ≤ 1, k ∈ K
where the primal variables zek = 1 if k ∈ Q and 0 otherwise.

Fixing the flow vector y , the subset Q ⊆ K that maximizes the traffic on
e ∈ E can be computed by linear programming:
X X
max dˆk yek zek = min ρke + πe Γ
k∈K k∈K
X
s.t. zek ≤Γ s.t. ρke + πe ≥ dˆk yek , k ∈ K
k∈K
0 ≤ zek ≤ 1, k ∈ K ρke , πe ≥ 0, k ∈ K,
where the dual variable πe ∼ GUB constraint and ρke ∼ zek ≤ 1.

Compact Γ-Robust Model
Compact Reformulaton:
X
min κe xe (4)
e∈E

X 1
 i = s(k)
k k
s.t. (yij − yji ) = −1 i = t(k) , ∀i ∈ V , k ∈ K (5)

{i,j}∈δ(i) 
0 else
X X
d¯k yek + Γπe + ρke ≤ Cxe , ∀e ∈ E (6)
k∈K k∈K
dˆk yek ≤ ρke + πe , ∀e ∈ E , k ∈ K (7)
y , x, ρ, π ≥ 0 (8)
x ∈ Z|E | (9)
Compared to the deterministic model, we have |E | + |E ||K | additional

variables and |E ||K | additional constraints.
How to solve RND?
DUALIZE = compact reformulation
SEPARATE = separation of robust inequalities

Outline

7 Back to practice

Cutset Polyhedra
Every cut in the network defines a two node problem
Consider the network design problem on a cut

The corresponding polyhedron is the cutset polyhedron PS
S V\S
d(S)

Cutset vs. Network Design Polyhedra
Definition 2
The cutset polyhedron PS for a given node-set S and the corresponding cut
δ(S) is defined as the network design polyhedron with the following data:
graph: GS := (VS , δ(S)) with nodes: VS := {S, V \ S} and links: δ(S)
demands: only those with endnodes in different
P shores:
QS := {k ∈ K : s(k) ∈ S, t(k) 6∈ S}, dS := k∈QS d k
S V\S One cutset-polyhedron for every cut

of the original network
d(S)
Theorem 3 (Raack, K., Orlowski, Wessäly [38])

Let S ⊂ V . Every non-trivial facet-defining inequality of PS defines a facet
of P if G [S] and G [V \ S] are (strongly) connected.
Cutset Inequalities
Let S ⊂ V and PS the corresponding cutset polyhedron.
S V\S
d(S)
capacity ≥ demand
l m
dS
C x(δ(S)) ≥ dS → x(δ(S)) ≥ C
Chvátal-Gomory cut (sum of capacity and flow conservation ineq.)

dS
Define facets of PS if C ∈
/Z
Define facets of P if in addition G [S] and G [V \ S] are strongly
connected.

The Robust-Γ Cutset Polyhedron
Subset S and the corresponding cutset δ(S)

QS ⊆ K all commodities with source s(k) ∈ S and target t(k) ∈ V \ S
Γ-Robust Cutset Polyhedron:
X
(yijk − yjik ) = 1 ∀k ∈ QS (10)
{i,j}∈δ(S)
X X
Γπe + d¯k yek + ρke ≤ Cxe ∀e ∈ δ(S) (11)
k∈QS k∈QS
ρke + πe − dˆk yek ≥ 0 ∀e ∈ δ(S), k ∈ QS (12)

x, y , ρ, π ≥ 0 (13)

Robust Cutset Inequalities
Aggregated mean cut-demand d¯S := k∈QS d¯k > 0

P
Aggregation of flow conservation (relaxing bwd flow):

X
d¯k y k (δ(S)) ≥ d¯S . (14)
k∈QS
Aggregation of capacity constraints:

X X
Γπ(δ(S)) + d¯k y k (δ(S)) + ρk (δ(S)) ≤ Cx(δ(S)) (15)
k∈QS k∈QS

(14) + (15):
X
Γπ(δ(S)) + d¯S + ρk (δ(S)) ≤ Cx(δ(S)) (16)
k∈QS
Consider Q ⊆ QS
Add all constraints ρke + πe − dˆk yek ≥ 0 for e ∈ δ(S) and k ∈ Q:
X
(Γ − |Q|)π(δ(S)) + d¯S + d(Q)
ˆ + ρk (δ(S)) ≤ Cx(δ(S)) (17)
k∈QS \Q
Add ρke ≥ 0 for all e ∈ δ(S) and k ∈ QS \Q results in:
Cx(δ(S)) + (|Q| − Γ)π(δ(S)) ≥ d¯S + d(Q)

ˆ (18)

Cx(δ(S)) + (|Q| − Γ)π(δ(S)) ≥ d¯S + d(Q)

ˆ
For which Q is the right hand side of this inequality the largest?
Among all Q with |Q| = i, the best choice are the i largest dˆk values
Let Qi be the set of i largest dˆk values
Let di := d¯S + d(Q
ˆ i)
For i = 0, . . . , |QS | we have
Cx(δ(S)) + (i − Γ)π(δ(S)) ≥ di . (19)
All these inequalities are linear combinations of model inequalities and thus
cannot be violated by LP solutions!

Mixed Integer Rounding:

Define r (d, c) := d − c( dc − 1)

With ri := r (di , C ), the MIR inequality reads

+ di
ri x(δ(S)) + (i − Γ) π(δ(S)) ≥ ri . (20)
C
For i = Γ this inequality reduces to the Γ-robust cutset inequality

&¯ ˆ Γ)
'
dΓ dS + d(Q
x(δ(S)) ≥ = (21)
C C

Theorem 4 (K., Kutschka, Raack [27])

Inequality (21)l defines
m a facet of PΓ (S) if and only if rΓ < C and either
dΓ
|δ(S)| = 1 or C ≥ 2.
One more MIR inequality might define a facet!
There is more!
Cx(δ(S)) + (i − Γ)π(δ(S)) ≥ di
To obtain further inequalities, we set x̄ := x(δ(S)) and π̄ := π(δ(S)) and
study the two dimensional set
X̄ = {(x̄, π̄) ∈ Z+ × R+ | C x̄ + (i − Γ)π̄ ≥ di for all i ∈ {0, . . . , |QS |}} .

Example: C = 5, Γ = 3, |QS | = 6, d¯S = 9, dˆ = (11, 8, 6, 6, 3, 1).
i=3 i=2 i=1 base, i=0
i=0
12 12
i=0, j=1
11 11
10 10
9 i=4
9 i=1,j=2
8 8
7 7
6 i=5 6
5 5
i=6
4 4
π(δ( S )) π(δ( S ))
3 i=4, j=5
3
2 2
1 1
MIR, base i=5
6 7 8 9 10 6 7 8 9 nonneg.10
x(δ (S)) x(δ (S))
Convex envelope: 5x(δ(S)) − 3π(δ(S)) ≥ 9, π(δ(S)) ≥ 0,x(δ(S)) ≥ 3,

2x(δ(S)) − π(δ(S)) ≥ 6, 3x(δ(S)) − π(δ(S)) ≥ 14, 7x(δ(S)) + 2π(δ(S)) ≥ 59,
3x(δ(S)) + 2π(δ(S)) ≥ 27
⇒ Complete description can be derived [27].
Separation strategies
robust cutset ineq. envelope ineq.

enum (ex) ILP (ex) shrinking (h) enum (ex) shrinking (h)
strategy (i) · · · · ·
strategy (ii) X · · · ·
strategy (iii) X · · X ·
strategy (iv) · · X · X
strategy (v) · X X · X
Considered strategies with exact (ex) and heuristic (h) separation algorithms

Separation of Cutting Planes
Additional gap closed in root for strategies (ii) and (iii)

Min, max, and geom. mean of additional gap closed in root (for all Γ)

Speed-up compared to CPLEX (strategy (i))

Outline

7 Back to practice

Price of Robustness
Price of Robustness = Extra cost compared to Γ = 0
Gain of Robustness = Cost savings compared with 95% deterministic model

Realized Robustness
Abilene network: Evaluation of 8064 traffic matrices (4 weeks)

realized robustness = maximal fraction of the total demand that can be
routed for a traffic matrix

Realized Robustness
Abilene network: Evaluation of 8064 traffic matrices (4 weeks)
Géant network: Evaluation of 2688 traffic matrices (4 weeks)
Γ A BILENE 1 A BILENE 2 G EANT 1 G EANT 2

0
1
5
10
|K|
Remarks
About 10% cost savings for Γ = 5 without losing robustness!
Speed-up of computations by generalized cutset inequalities
Further studies on impact of choice of d¯k and dˆk [26]

Part III
Extensions to Robust Optimization

Robust Optimization Revisited
Advantages Robust Optimization:

Only marginal complexity increase (compared to deterministic case)
Trade-off between level of robustness and cost of solution by parameter Γ
Optimizes result in the worst-case (in advance)
No information on probability distribution needed
Disadvantages Robust Optimization:

Single solution without any flexibity!
The almost always optimal solution might be infeasible
⇒ Two-Stage Robustness Concepts
Very inprecise description of uncertainty (only two values)
⇒ More detailed description of uncertainty

Outline
8 Two-Stage Robust Optimization

9 Robust Network Design with Affine Recourse
10 Recoverable Robustness

Classical Network Design Model
Integer Linear Programming formulation:
X
min κe xa
a∈A
X X
s.t. yak − yak = dk ∀k ∈ K
a∈δ + (s k ) a∈δ − (s k )
X X
yak − yak =0 ∀k ∈ K , ∀i ∈ V , i 6= s k , t k
a∈δ + (i) a∈δ − (i)
X
C xa − yak ≥0 ∀a ∈ A
k∈K
xa ∈ Z+ , yak ≥ 0
Uncertain demand values d k in right hand side

For robust solution, worst case d k value have to be taken, say d¯k + dˆk
Can’t we do better! What about dynamic routing?

Dynamic Routing Model
For d ∈ U, yak (d) denotes the flow on arc a.
X
min κa xa
a∈A
X X
s.t. yak (d) − yak (d) = d k (d) ∀d ∈ U, k ∈ K
a∈δ + (s k ) a∈δ − (s k )
X X
yak (d) − yak (d) =0 ∀d ∈ U, k ∈ K , i ∈ V \ {s k , t
a∈δ + (i) a∈δ − (i)
X
C xa − yak (d) ≥0 ∀d ∈ U, a ∈ A
k∈K
xa ∈ Z+ , yak (d) ≥ 0
Extremely large model!

Projection on Capacity Space
Reduction model size by projection on capacity space

For fixed demand:
P x = projx P
|E |
= conv{x ∈ Z+ : ∃ f ∈ [0, 1]|P| such that (f , x) ∈ P}
|E |
= conv{x ∈ Z+ : x satisfies (22) for all `M ∈ Met(G )}
where X X
`M (e)xe ≥ d k `M (s k , t k ), (22)
e∈E k∈K
and Met(G ) denotes the cone of metrics in G .
Note, for static / stable routing this result does not hold!

Metric Inequalities for Dynamic Routing
Metric Inequalities for Dynamic Routing [31]

X X
`M (e)xe ≥ max d k `M (s k , t k ), (23)
d∈U
e∈E k∈K
Theorem 5 (Mattia, 2013)

|E |
P x (U) = conv{x ∈ Z+ : x satisfies (23) for all `M ∈ Met(G )}
Theorem 6 (Mattia, 2013)

If ax ≥ b is valid for P x (U), then there exists a metric `M ∈ Met(G ) such
that `M x ≥ b is still valid and `M (e) ≤ ae for all e ∈ E .

Separation of Metric Inequalities
Separation of Metric Inequalities without Uncertainties: I.e., check whether

x ? satisfies all metric inequalities (not tight).
Let α be the percentage of all aggregated commodities that can be routed
with capacities x ? .
max α
X
s.t. (yijk − yjik ) = αdvk ∀v ∈ V , k ∈ K [µkv ]
{i,j}∈δ(v )
X
yek ≤ xe? ∀e ∈ E [`(e)]
k∈K
y ≥0
If α ≥ 1, x ? satisfies all metric inequalities (22).

Otherwise, there exists a violated metric inequality.

Separation of Metric Inequalities
Separation of Metric Inequalities without Uncertainties:
X XX
z ? = min xe? `(e) − dvk µkv
e∈E k∈K v ∈V
s.t. µkv − µkw ≤ `(e) ∀k ∈ K , vw ∈ E
µkw − µkv ≤ `(e) ∀k ∈ K , vw ∈ E
XX
dvk µkv = 1
k∈K v ∈V
`(e) ≥ 0
If z ? ≥ 01, x ? satisfies all metric inequalities (22).
Otherwise, X XX
`(e)? xe ≥ dvk µk?
v 1
e∈E k∈K v ∈V
is a violated metric inequality.
Note, µkv denotes length of shortest path from s k to v
Separation of Dynamic Metric Inequalities
Observation: Only right-hand-side is different, but requires a bilevel
formulation:
X XX
z ? = min xe? `(e) − max dvk (d)µkv β
d∈U
e∈E k∈K v ∈V
s.t. µkv − µkw ≤ `(e) ∀k ∈ K , vw ∈ E
µkw − µkv ≤ `(e) ∀k ∈ K , vw ∈ E
X
`(e) = 1
e∈E
`(e) ≥ 0
with
( )
XX
β = max dvk (d)µkv : Ad ≤ b, d ≥ 0
k∈K v ∈V
Let U = {d ≥ 0 : Ad ≤ b} be a polyhedral uncertainty set.

Summary Dynamic Routing
Dynamic Routing is just one example of two-stage robust optimization:

At the first stage, some decisions have to be made before the uncertain
data is known
Af the second stage, the uncertain data becomes available and the
remaining decisions have to be taken.
Two-stage Robust Optimization vs. Two-stage Stochastic Optimization

T T
" # " #
min q z min q z
min c T x + max Bz = h − Tx minc T x + E Bz = h − Tx
z ≥0 z ≥0
s.t. Ax = b s.t. Ax = b
x ≥0 x ≥0

Robust Maximum Flow Revisited
Maximum Flow in directed graph G = (V , A) with uncertain capacities ua
What flow value z can be guaranteed without fixing the flow?
1st Stage: Fix z independent of realization u ∈ U

2nd Stage: Find flow x of value at least z for every u ∈ U
max z

X X z
 v =s
s.t. xa (u) − xa (u) = −z v =t ∀u ∈ U, v ∈ V

a∈δ + (v ) a∈δ − (v ) 
0 otherwise
xa (u) ≤ ua (u) ∀u ∈ U, a ∈ A
z ≥ 0, xa (u) ≥ 0
Note: 2nd stage does not mean this is done afterwards; optimization
integrates both stages!
Robust Maximum Flow
Theorem 7 (Minoux, 2010)

The robust maximum flow problem is solvable in polynomial time
for a fixed number of K scenarios
for U = [ū1 − û1 , ū1 ] × [ū2 − û2 , ū2 ] × · · · × [ūm − ûm , ūm ] (with
A = {1, . . . , m}), i.e., interval scenarios
The problem is strongly NP-hard if U is defined budget uncertainty
(Γ-robustness)
The latter problem is polynomial time solvable if G is planar.
The results [32] can be extended to general LPs with uncertain
right-hand-side [33].

Outline


Static vs. Dynamic Routing
Static Routing:
Capacities have to be installed in integer amounts
Routing templates fixes percentual distribution of traffic volume along
paths
Dynamic Routing:
Routing can be adapted to actual traffic volumes (realization from
uncertainty set)
Compromise:
Routing “follows” size of demands: affine routing

Affine Routing Function
Robust Network Design with Affine Routing:

Routing follows a linear function of all traffic values
X
yijk (d) := hijk0 + hijk k̄ d k̄
k̄∈K
where hijk0 , hijk k̄ ∈ R for all ij ∈ A, k, k̄ ∈ K .
Theorem (Poss & Raack [37])

Let D be an arbitrary demand uncertainty set. Then
OPTdyn (D) ≤ OPTaff (D) ≤ OPTstat (D)

Outline


Recoverable Robustness
Recoverable robustness [28, 13]

uncertainty as two-stage process:
1st stage: a-priori decision
2nd stage: recovery:
limited change of first-stage decision
after realization of uncertainty is known
optimize worst-case w. r. t. recovery
Example:
Recoverable Robust Knapsack problem (RRKP) with
Discrete Scenarios [15]
Γ Scenarios [16]
Recoverable Robust Network Topology Design (discrete scenarios) [3]

(k, `)-RRKP with Discrete Scenarios
Given items N = {1, . . . , n},
first stage: profits p 0 , weight w 0 , capacity c 0 ,
scenarios S ∈ SD with profits p S , weight w S , capacity c S ,
recovery set X (X ): delete ≤ k items, add ≤ ` items
Find subset X ⊆ N
Such that w 0 (X ) ≤ c 0 ,
for all S ∈ SD there exists X S ∈ X (X ) with w S (X S ) ≤ c S ,
total profit
pT (X ) = p 0 (X ) + min max p S (X S )
S∈SD X S
is maximized.
First S_1 S_2

RRKP with Discrete Scenarios - ILP
X
max pi0 xi + ω
i∈N
X
s. t. wi0 xi ≤c 0 first stage second stage rem
i∈N addition of ≤ ` items
X
wiS xiS ≤c S ∀S ∈ SD
i∈N
xi − xiS − yiS ≤0 ∀S ∈ SD , i ∈ N
X n
yiS ≤k ∀S ∈ SD
i=1
xiS − xi − ziS ≤0 ∀S ∈ SD , i ∈ N
X n
ziS ≤` ∀S ∈ SD
i=1
X n
ω− wiS xiS ≤0 ∀S ∈ SD
i=1
xi , xi , yiS , ziS ∈
S
{0, 1}

k-RRKP with Γ Scenarios
Given Items N = {1, . . . , n},
first stage: profits p 0 , weights w 0 , capacity c 0 ,
Γ-scenarios: weights [w̄ , w̄ + ŵ ], capacity c, Γ ∈ N,
recovery set X (X ): delete ≤ k items from X ⊆ N
Find subset X ⊆ N,
Such that w 0 (X ) ≤ c 0 ,
for all S ∈ SΓ there exists X S ∈ X (X ) with w S (X S ) ≤ c,
total profit p 0 (X ) is maximized
First Second S_1

RRKP with Γ scenarios - MP
Mathematical Programming formulation:

X
max pi0 xi
i∈N
X
s. t. wi0 xi ≤c 0
i∈N
 !
X X X X
w̄i xi + max  ŵi xi − max w̄i xi + ŵi xi  ≤c
X ⊆N Y ⊆N
i∈N |X |≤Γ i∈X |Y |≤k i∈Y i∈X ∩Y
xi ∈ {0, 1}
Question: Compact Linear reformulation?

Answer: LP duality and enumeration of solution values!

RRKP with Γ scenarios - ILP
Let U := {0} ∪ {w̄i : i ∈ N} ∪ {w̄i + ŵi : i ∈ N}

Then, a compact reformulation is:
X
max pi0 xi
i∈N
X
s. t. wi0 xi ≤c 0
i∈N
X X X
w̄i xi + uxi + Γξ u + θiu ≤c + ku ∀u ∈ U
i∈N: i∈N: i∈N
w̄i <u w̄i ≥u
min{−w̄i + u, ŵi }xi − ξ u − θiu ≤0 ∀u ∈ U

u
ξ , θiu ≥0 ∀u ∈ U, i ∈ N
Resulting compact model contains O(n2 ) variables and constraints

Complexity of Robust Knapsack
Theorem 8 (Karp 72, Bellman 57)

The knapsack problem is weakly NP-hard, i.e., it can be solved in O(nc)
time.
Theorem 9 (Yu 96, Kalai & Vanderpooten 06)

The robust knapsack problem with bounded number of scenarios can be
solved in pseudo-polynomial time.
Theorem 10 (Yu 96, Aissi et al. 07)

The robust knapsack problem with discrete scenarios is strongly NP-hard
and not approximable, unless P = NP.
Theorem 11 (Bertsimas & Sim [11, 12], Klopfenstein & Nace [23])
The Γ-robust knapsack problem can be solved in pseudo-polynomial time.

Complexity of RRKP with Discrete Scenarios
Theorem 12 (Büsing, K., Kutschka [15])

The (k, `)-rrKP is strongly NP-hard for unbounded sets of discrete
scenarios even if either p 0 = 0 or p S = 0 for all S ∈ SD holds.
Reductions from 3SAT.
Corollary 13 (Büsing, K., Kutschka [15])
The (k, `)-rrKP cannot be approximated within `+1 ` , unless P = NP. In
particular, for ` = 0, the problem cannot be approximated.

The (k, `)-rrKP can be solved in pseudo-polynomial time for a bounded
number of scenarios.
Generalization of dynamic programming for robust knapsack (Yu, 1996).

Complexity of RRKP with Γ Scenarios

The RRKP with Γ scenarios is at least weakly NP-hard.
Open Problem
Is the RRKP with Γ scenarios strongly NP-hard or does there exist a
pseudo-polynomial time algorithm?

Part IV
Conclusions

Concluding Remarks
Robust optimization is an emerging field in mathematical optimization

but requires an additional effort to solve robust problems
Correct modelling of uncertainties in network applications required
Integration of other (real) networking aspects (survivability, multi-layer)
⇒ Real applications [7, 20, 19]
⇒ Robustness & 1+1 Protection [24]
Recoverable Robustness allows a two-stage approach
⇒ What recovery action is possible in Network Design?
Quality of robust approach has to be evaluated
⇒ Which value of Γ is enough to obtain robust designs?

Part V
Exercises & Research Questions

Exercise 1
Consider the Γ-robust knapsack problem

( n n
)
X X
max ci xi : ai xi ≤ b, xi ∈ {0, 1}
i=1 i=1
 ai ∈ [ā
where ai are random variables, i − âi 
, āi + âi ].  
50 41 18
30  47   28 
     
Let n = 5, b = 250, c =  45, ā =  64 , and â =  45 .
    
25  22   14 
70 113 103
Determine the optimal solution for Γ = 0, 1, 2, 3, 4, and 5.
Hint: If you are not sure you found the optimal solution, try exercise 3 first.

Exercise 2
Let A be a m × n matrix. Consider the following uncertain linear

optimization problem:
min{c T x : Ax ≤ b},
x
under the uncertainty:
U = {(c, A, b) : |cj − c̄j | ≤ σj , Aij − Āij ≤ αij , bi − b̄i ≤ βi , ∀i, j},
where c̄j , etc. denotes the nominal data.

Reduce the robust counterpart of the problem to a linear program with
m constraints (not counting the non-negativity constraints) and
2n nonnegative variables.

Exercise 3
The knapsack problem max{c T x : aT x ≤ b, x ∈ {0, 1}n } can be solved by

a dynamic programming algorithm in O(nb) time. For this, the function
( k k
)
X X
f (k, d) := max ci xi : ai xi = d, xi ∈ {0, 1}
i=1 i=1
is solved for all k ∈ {0, . . . , n} and d ∈ {0, . . . , b}.

Develop a dynamic programming algorithm for the Γ-robust knapsack
problem. What is the running time?

Exercise 4
Consider again the robust knapsack problem, but this time with a
polyhedral uncertainty set
U = a ∈ Rn+ : Da ≤ d

where D has k rows and n columns and d is a k-dimensional vector.

Is it possible to generalize the dynamic programming algorithm for the
knapsack problem to this robust version?
If yes, what is the running time?
If no, how can we prove that the problem is strongly NP-complete?

Exercise 5
For the knapsack problem, a cover is a subset C ⊆ {1, . . . , n} of the items

such that a(C ) > b. The cover is minimal if a(C \ {i}) ≤ b for all i ∈ C .
The valid cover inequality
X
xi ≤ |C | − 1
i∈C
defines a facet of the knapsack problem if C is minimal and |C | = n. In

general, the inequality can be strengthened by extending the set C or lifting
items outside C .
Find a minimal cover for the example of exercise 1 with a = ā
(deterministic case, Γ = 0).
How does the definition of a cover change in case of the Γ-robust
knapsack problem?

Exercise 6
Combined-Heat-and-Power Day-Ahead Planning:
qt heat
ft storage
fuel CHP
ut
zt dth heat
demand
ptg
power ptb dtp power

$ market demand
pts
Time horizon T (T = 24), demand for power dtp and heat dth
Fuel ft to operate CHP costs ctf per unit
Power bought ptb /sold pts on day-ahead market at ctp per unit
Generation at time t: ptg (power), qt (heat), zt (on/off)
Fixed ratio ρ between heat and power generation
Min. / Max. heat production bounds: Q, Q̄
Heat can be stored with loss factor α per time unit
Storage at end of period t: ut , lower and upper bounds U, Ū
Model this planning problem as MILP
Exercise 6
Heat demand is uncertain:

Discrete Uncertainty Set U = {d h,k : k = 1, . . . , K }
Model Robust CHP Planning Problem

Part VI
References

References I
[1] A. Altin, E. Amaldi, P. Belotti, and M. Ç. Pinar. Provisioning virtual private networks under traffic
uncertainty. Networks, 49(1):100–155, 2007.
[2] A. Altin, H. Yaman, and M. C. Pinar. The Network Loading Problem under Hose Demand Uncertainty:
Formulation, Polyhedral Analysis, and Computations. INFORMS Journal on Computing, 23:75–89, 2011.
[3] E. Alvarez-Miranda, I. Ljubic, S. Raghavan, and P. Toth. The recoverable robust two-level network design
problem. INFORMS Journal on Computing, 27(1):1–19, 2014.
[4] A. Atamtürk and M. Zhang. Two-Stage Robust Network Flow and Design Under Demand Uncertainty.
Operations Research, 55(4):662–673, 2007.
[5] P. Avella, S. Mattia, and A. Sassano. Metric inequalities and the network loading problem. Discrete
Optimization, 4:103–114, 2006.
[6] P. Belotti, A. Capone, G. Carello, and F. Malucelli. Multi-layer MPLS network design: The impact of
statistical multiplexing. Computer Networks, 52(6):1291–1307, 2008.
[7] P. Belotti, K. Kompella, and L. Noronha. A comparison of OTN and MPLS networks under traffic
uncertainty. Working paper, 2011.
http://myweb.clemson.edu/~pbelott/papers/\robust-opt-network-design.pdf.
[8] W. Ben-Ameur. Between fully dynamic routing and robust stable routing. In Design and Reliable
Communication Networks, 2007. DRCN 2007. 6th International Workshop on, pages 1–6. IEEE, 2007.
[9] W. Ben-Ameur and H. Kerivin. Routing of uncertain demands. Optimization and Engineering, 3:283—313,
2005.
[10] A. Ben-Tal, L. E. Ghaoui, and A. Nemirovski. Robust optimization. Princeton University Press, 2009.
[11] D. Bertsimas and M. Sim. Robust discrete optimization and network flows. Math. Program., Ser. B
98:49–71, 2003.

References II
[12] D. Bertsimas and M. Sim. The Price of Robustness. Operations Research, 52(1):35–53, 2004.
[13] C. Büsing. Recoverable Robustness in Combinatorial Optimization. PhD thesis, Technische Universität
Berlin, 2011.
[14] C. Büsing. Recoverable robust shortest path problems. Networks, 59(1):181–189, 2012.
[15] C. Büsing, A. M. C. A. Koster, and M. Kutschka. Recoverable robust knapsacks: the discrete scenario case.
Optimization Letters, 5(3):379–392, 2011.
[16] C. Büsing, A. M. C. A. Koster, and M. Kutschka. Recoverable robust knapsacks: gamma-scenarios. EURO
Journal on Computational Optimization, 7(1):15–45, 2019.
[17] F. D. C. Büsing. New results about multi-band uncertainty in robust optimization. Proceedings of SEA 2012,
11th Symposium on Experimental Algorithms, Lecture Notes on Computer Science 7276, pages 63–74, 2012.
[18] G. Claßen, A. M. C. A. Koster, M. Kutschka, and I. Tahiri. Robust Metric Inequalities for the Γ-Robust
Network Loading Problem. Networks, submitted.
[19] G. Claßen, A. M. C. A. Koster, and A. Schmeink. A robust optimisation model and cutting planes for the
planning of energy-efficient wireless networks. Computers & Operations Research, 40(1):80–90, 2013.
[20] S. Duhovniko, A. M. C. A. Koster, M. Kutschka, F. Rambach, and D. Schupke. Γ-robust network design for
mixed-line-rate-planning of optical networks. In Proc. Optical Fiber Communication - National Fiber Optic
Engineers Conference (OFC/NFOEC), 2013.
[21] M. Fischetti and M. Monaci. Cutting plane versus compact formulations for uncertain (integer) linear
programs. Mathematical Programming Computation, 4:239–273, 2012.
[22] M. Iri. On an Extension of the Maximum-flow Minimum-cut Theorem to Multicommodity Flows. Journal of
the Operations Research Society of Japan, 13(3):129–135, 1971.

References III
[23] O. Klopfenstein and D. Nace. Cover inequalities for robust knapsack sets – application to the robust
bandwidth packing problem. Networks, 59(1):59–72, 2012.
[24] A. M. C. A. Koster and M. Kutschka. An integrated model for survivable network design under demand
uncertainty. In Proceedings of 8th International Workshop on the Design of Reliable Communication
Networks (DRCN 2011), pages 54–61, 2011.
[25] A. M. C. A. Koster, M. Kutschka, and C. Raack. Towards robust network design using integer linear
programming techniques. In Proceedings Next Generation Internet (NGI), 2010.
[26] A. M. C. A. Koster, M. Kutschka, and C. Raack. On the robustness of optimal network designs. In
Proceedings IEEE International Conference on Communications (ICC), 2011.
[27] A. M. C. A. Koster, M. Kutschka, and C. Raack. Robust Network Design: Formulations, Valid Inequalities,
and Computations. Networks, 61(2):128–149, 2013.
[28] C. Liebchen, M. E. Lübbecke, R. H. Möhring, and S. Stiller. The concept of recoverable robustness, linear
programming recovery, and railway applications. In R. Ahuja, R. Möhring, and C. Zaroliagis, editors, Robust
and Online Large-Scale Optimization, volume 5868 of Lecture Notes on Computer Science, pages 1–27, 2009.
[29] M. Lomonosov. Combinatorial approaches to multiflow problems. Discrete Applied Mathematics, 11:1–93,
1985.
[30] T. L. Magnanti, P. Mirchandani, and R. Vachani. The convex hull of two core capacitated network design
problems. Mathematical Programming, 60:233–250, 1993.
[31] S. Mattia. The robust network loading problem with dynamic routing. Computational Optimization and
Applications, 54(3):619–643, 2013.
[32] M. Minoux. On robust maximum flow with polyhedral uncertainty sets. Optimization Letters, 3:367–376,
2009.

References IV
[33] M. Minoux. On 2-stage robust LP with RHS uncertainty: complexity results and applications. Journal on
Global Optimization, 49:521–537, 2011.
[34] M. Minoux. Two-stage robust LP with ellipsoidal right-hand side uncertainty is NP-hard. Optimization
Letters, 6:1463–1475, 2012.
[35] M. Monaci, U. Pferschy, and P. Serafini. Exact solution of the robust knapsack problem. Computers &
Operations Research, 40(11):2625–2631, 2013.
[36] K. Onaga and O. Kakusho. On feasibility conditions of multicommodity flows in networks. Circuit Theory,
IEEE Transactions on, 18(4):425 – 429, July 1971.
[37] M. Poss and C. Raack. Affine recourse for the robust network design problem: between static and dynamic
routing. Networks, 61(2):180–198, 2013.
[38] C. Raack, A. M. C. A. Koster, S. Orlowski, and R. Wessälly. On cut-based inequalities for capacitated
network design polyhedra. Networks, 57:141–156, 2011.
[39] D. Rajan and A. Atamtürk. On splittable and unsplittable flow capacitated network design arc-set polyhedra.
Mathematical Programming, 92:315–333, 2002.

NetOpt2020 2020 01 20 Arie

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

NetOpt2020 2020 01 20 Arie

Uploaded by

Copyright:

Available Formats

Robust Network Optimization

Arie M.C.A. Koster

9th Winter School on Network Optimization – 20 January 2020

Introduction to Robust Optimization (1st lecture)

Arie Koster – RWTH Aachen University 2 / 101

Arie Koster – RWTH Aachen University 3 / 101

Arie Koster – RWTH Aachen University 4 / 101

Network Design: Capacity Installation, Routing, Survivability

Demand uncertainties Traffic fluctuations in the US abilene Internet2

Multi-period Network Design

Arie Koster – RWTH Aachen University 7 / 101

Heat can be stored for future use, power cannot be stored

Power has to be bought/sold at

Arie Koster – RWTH Aachen University 8 / 101

Arie Koster – RWTH Aachen University 9 / 101

Chance-constrained Optimization Problem (COP)

How to solve a COP?

Arie Koster – RWTH Aachen University 10 / 101

Chance-Constrained Linear Programming with joint constraintswith

with Entries of A, b and/or c are not constant but random variables

Arie Koster – RWTH Aachen University 11 / 101

How to solve this problem?

Arie Koster – RWTH Aachen University 12 / 101

Arie Koster – RWTH Aachen University 13 / 101

After relaxing the integrality of x, a second order cone problem remains,

Arie Koster – RWTH Aachen University 15 / 101

Uncertain Linear Program

where all input data stems from an uncertainty set U ⊂ Rn × Rm×n × Rm .

Arie Koster – RWTH Aachen University 16 / 101

Robust feasible solution

Robust solution value

min {ĉ(x) : x is robust feasible}

The robust counterpart can be written as

min{c T x : (Ā + Â)x ≤ b̄, x ≥ 0}

Arie Koster – RWTH Aachen University 18 / 101

Note: the robust counterpart does not change if Û = U1 × U2 × . . . × Um

Arie Koster – RWTH Aachen University 19 / 101

How to define the uncertainty sets?

Ui = (a, b) ∈ Rn+1 : k(a, b) − (ā, b̄)k < κ

Uncertainty set is an polyhedron, e.g.,

Ui = (a, b) ∈ Rn+1 : D · (a, b) ≤ d

with D ∈ Rk×n , d ∈ Rk for some k ∈ N [2].

Arie Koster – RWTH Aachen University 20 / 101

Arie Koster – RWTH Aachen University 21 / 101

Ui (Γ) = {ai ∈ Rn : aij = āij + âij zij ∀j = 1, . . . , n, zi ∈ Zi (Γ)}

Arie Koster – RWTH Aachen University 23 / 101

Arie Koster – RWTH Aachen University 24 / 101

Theorem 1 (Bertsimas & Sim [12])

Arie Koster – RWTH Aachen University 25 / 101

Choice of Γ as a function of n so that the

Note: Result is independent of actual distribution of random variables aij ,

Arie Koster – RWTH Aachen University 26 / 101

Arie Koster – RWTH Aachen University 27 / 101

This number of inequalities is exponential if Γ = O(n)

Arie Koster – RWTH Aachen University 28 / 101

Given x ? , βi (x ? , Γ) is the optimization problem

Arie Koster – RWTH Aachen University 30 / 101

Thus, (1) now reads

Arie Koster – RWTH Aachen University 31 / 101

Network Design under Demand

Arie Koster – RWTH Aachen University 32 / 101

5 Robust Network Design

Arie Koster – RWTH Aachen University 33 / 101