Professional Documents
Culture Documents
2-Lewis Ntaimo
2-Lewis Ntaimo
Lewis Ntaimo
Michelle M. Alvarado and Guglielmo Lulli†
1 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Agenda
Methodology
Subgradient Decomposition
Fenchel Decomposition (FD)
Fenchel Disjunctive Decomposition (FDD)
Example Illustration
Summary
2 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
n0
x ∈ Zn̄+1 × R+1 is a decision variable vector
λ > 0 is a suitable weight factor
n0
X ⊆ Rm1 is a nonempty polyhedron, X = {Ax ≥ b, x ∈ Zn̄+1 × R+1 }
ω̃ multivariate random variable and for a given x ∈ X the real random cost
variable is given as: f (x, ω̃) = c > x + Q(x, ω̃)
3 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
More Notation:
The family of real random cost variables {f (x, ω̃)}x∈X ⊆ F are defined on a
probability space (Ω, F, P)
E : F 7→ R denotes the expected value
D : F 7→ R is a mean-risk measure
F is the space of all real random cost variables f : Ω 7→ R
A scenario defines the realization of the stochastic problem data
{h(ω), T (ω)}
I W ∈ Rm2 ×n2 is a fixed recourse matrix
Assumptions:
(A1) The random variable ω̃ is discrete with finitely many scenarios ω ∈ Ω, each
with probability of occurrence p(ω).
(A2) The first-stage feasible set is nonempty, that is, X 6= ∅.
n0
(A3) The second-stage feasible set {Wy (ω) ≥ h(ω) − T (ω)x, y (ω) ∈ Zn̄+2 × R+2 }
is bounded and nonempty for all x ∈ X .
4 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Mean-Risk Measures
Expected Excess (EE): EE is the expected value of the excess over the target
η ∈ R.
D = φEEη (x) = E[max{f (x, ω̃) − η, 0}].
Absolute Semideviation (ASD): ASD reflects the expected value of the excess
over the mean value.
5 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Motivation:
Limited algorithms available to solve this class of problems despite many
applications
Example approach: Dual decomposition based (Schultz, 2011)
Applications of interest: chemotherapy scheduling, wildfire risk
management, air traffic flow management
Properties and challenges:
Large-scale nature, mean-risk objective function is lower
semi-continuous, generally nonconvex and possibly discontinuous
(Schultz, 2011)
Decomposing mean-risk SMIP based on semideviation scenario-wise
does not result in separable scenario problems
6 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Related Work
SMIP with Mean-Risk Description Decision Vars.
(Märkert and Schultz, 2005) Deviation measures mixed-integer
(Schultz and Tiedemann, 2006) CVAR mixed-integer
(Miller and Ruszczyn’ski, 2009) QDEV continuous
(Schultz, 2011) Excess prob., CVAR, EE, mixed-integer
and ASD
Relevant Methodology
(Carøe, 1998) Disjunctive cuts integer
(Carøe and Schultz, 1999) Dual decomposition integer
(Sen and Higle, 2005) Disjunctive decomposition mixed-binary
(Ntaimo, 2013) Fenchel decomposition (FD) mixed-integer
(Ahmed, 2006) Subgradient decomposition continuous
(Cotton and Ntaimo, 2013) Subgradient decomposition continuous
7 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Methodology
Subgradient Decomposition
I Stage-wise decomposition: x, y (ω)
I Generate subgradients to sequentially
approximate convex E and D terms
Fenchel Decomposition
I Generate FD cuts to cut off fractional
LP-relaxation solution
I FD cuts are in y (ω)-space or (x, y (ω))-space
I Use disjunctive programming to lift cuts to
(x, y (ω))-space
Subgradient Decomposition
Master Problem:
Min (1 − λ)c > x k + (1 − λ)θ + λζ
s.t. ψτ> x + θ ≥ ψτ0 , τ = 1, · · · , k
(4)
στ> x + ζ ≥ στ0 , τ = 1, · · · , k
x ∈ X , θ, ζ free
Subgradient Decomposition
11 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
12 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
c (ω, x)
Convex hull of YIP (ω, x): YIP
c (ω)
Convex hull of YIP (ω): YIP
13 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Theorem
Let ω̄ denote a scenario with a fractional
subproblem LP-relaxation solution ŷ (ω̄) for a
given x̂ ∈ X . Define
g (ω̄, x̂, β) = Max {β > y | y ∈ YIP c (ω̄, x̂)} and
14 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
15 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Author's personal
Example copy
Illustration Fenchel Disjunctive Decomposition (FDD)
Author's personal copy
Summary
160 J Glob Optim (2013) 55:141–163
Sample results for risk neutral case: FD and FD-L algorithms
160 J Glob Optim (2013) 55:141–163
Table 9 Computational Results on SSLP using FD
Table 9 ComputationalFD
Instance Results on SSLP using FD
Algorithm CPU (s)
Instance FD
FD Algorithm
Iters FD cuts L2 cuts CPU
Min (s) Max Avg
SSLP5.25.50 FD2.80
Iters FD cuts
3.40 L2 cuts
0.00 Min0.37 Max0.69 Avg0.54
SSLP5.25.100
SSLP5.25.50 6.40
2.80 15.00
3.40 0.20
0.00 0.52
0.37 3.19
0.69 1.24
0.54
SSLP10.50.50
SSLP5.25.100 30.60
6.40 495.60
15.00 1.00
0.20 40.31
0.52 79.02
3.19 59.50
1.24
SSLP10.50.100
SSLP10.50.50 32.40
30.60 936.60
495.60 1.00
1.00 78.39
40.31 134.47
79.02 102.75
59.50
SSLP10.50.500
SSLP10.50.100 28.40
32.40 4669.60
936.60 1.00
1.00 376.54
78.39 517.81
134.47 451.80
102.75
SSLP15.45.5
SSLP10.50.500 94.60
28.40 321.20
4669.60 1.00
1.00 7.99
376.54 526.85
517.81 136.67
451.80
SSLP15.45.10
SSLP15.45.5 77.75
94.60 588.00
321.20 1.00
1.00 17.58
7.99 376.75
526.85 171.45
136.67
SSLP15.45.15
SSLP15.45.10 118.20
77.75 1318.40
588.00 1.00
1.00 22.61
17.58 771.80
376.75 338.37
171.45
SSLP15.45.20
SSLP15.45.15 78.40
118.20 1139.20
1318.40 1.00
1.00 100.06
22.61 300.25
771.80 166.72
338.37
SSLP15.45.25
SSLP15.45.20 83.50
78.40 1393.00
1139.20 1.00
1.00 85.41
100.06 244.59
300.25 168.69
166.72
SSLP15.45.25 83.50 1393.00 1.00 85.41 244.59 168.69
Table 10 Computational results on SSLP using FD-L
Table 10 Computational
Instance FD-Lresults on SSLP using FD-L
Algorithm CPU (s)
Instance FD-L
FD-L Algorithm
Iters FD-L cuts L2 cuts CPU
Min (s) Max Avg
FD-L Iters FD-L cuts L2 cuts Min Max Avg
SSLP5.25.50 2.40 2.80 0.00 0.36 0.70 0.52
SSLP5.25.100
SSLP5.25.50 6.00
2.40 12.80
2.80 0.00
0.00 0.37
0.36 2.56
0.70 1.10
0.52
SSLP10.50.50
SSLP5.25.100 31.40
6.00 503.20
12.80 1.00
0.00 37.51
0.37 75.79
2.56 57.74
1.10
SSLP10.50.50
SSLP10.50.100 31.40
31.20 503.20
899.40 1.00
1.00 37.51
69.44 75.79
108.48 57.74
91.57
SSLP10.50.100
SSLP10.50.500 31.20
28.60 899.40
4601.60 1.00
1.00 69.44
345.20 108.48
476.56 91.57
416.07
SSLP10.50.500
SSLP15.45.5 28.60
94.80 4601.60
327.20 1.00
1.00 345.20
7.83 476.56
553.60 416.07
138.61
SSLP15.45.5
SSLP15.45.10 94.80
84.25 327.20
637.50 1.00
1.00 7.83
27.39 553.60
404.66 138.61
181.53
SSLP15.45.10
SSLP15.45.15 84.25
117.00 637.50
1306.60 1.00
1.00 27.39
21.58 404.66
723.06 181.53
333.20
SSLP15.45.15
SSLP15.45.20 117.00
80.60 1306.60
1164.00 1.00
1.00 21.58
102.71 723.06
247.63 333.20
166.69
SSLP15.45.20
SSLP15.45.25 80.60
83.25 1164.00
1383.00 1.00
1.00 102.71
82.57 247.63
201.81 166.69
151.61
SSLP15.45.25 83.25 1383.00 1.00 82.57 201.81 151.61
on the first five problems. However, FD and FD-L perform better on the last five problems,
on the first
which havefive problems.
larger However,
first-stage FDdimension
decision and FD-Lspace.
perform betterthe
Unlike onFD
the and
last D2
five algorithms,
problems, 16 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
17 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Proposition
Suppose we have binary x̂ ∈ X . Then
c
conv {y | (x̂, y ) ∈ YIP (ω)} = {y | (x̂, y ) ∈ YIP (ω)}
(Carøe, 1998)
c
This result relates facets of {y | (x̂, y ) ∈ YIP (ω)} to
conv {y | (x̂, y (ω)) ∈ YIP (ω)} for binary x̂ ∈ X .
This implies that we can lift facets (valid inequalities) of the form
β > y ≤ g (ω, x̂, β) into the (x, y (ω))-space.
To accomplish this, we use result by Araoz et al. (1984) on next slide
applied to our setting
18 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Proposition
The inequality α(ω̄)> x + ϕβ > y ≤ ϕg (ω̄, x̂, β) + α(ω̄)> x̂, where ϕ ≥ 0, is a facet
extension of β > y ≤ g (ω̄, x̂, β) if and only if ϕ > 0 and α(ω̄)/ϕ is a vertex of the
polyhedron
This result allows for β > y ≤ g (ω̄, x̂, β) to be lifted to a valid inequality for
c
YIP (ω̄) if the extreme points of (6) are known.
c
We want valid inequalities for YIP (ω) for each ω ∈ Ω.
19 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Disjunctive programming:
Let i be an index for a binary component of z = (x, y ) and define
R = {0, 1}.
S0 (ω) = YLP (ω) ∩ {−zi ≥ 0} and S1 (ω) = YLP (ω) ∩ {zi ≥ 1}, where
Sr (ω) 6= ∅ for all r ∈ R.
disjunction of the two sets: S(ω) = ∪r ∈R Sr (ω)
c c
Since YIP (ω) ⊆ S(ω), valid inequalities for S(ω) are also valid for YIP (ω)
c
So our aim is to derive valid inequalities for YIP (ω) via S(ω) using
disjunctive programming
I Motivation: Generating valid inequalities for S(ω) involves solving LPs
instead of a subgradient optimization problem
20 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Definition
The reverse polar of the set S(ω), denoted S ] (ω), is defined as follows:
S ] (ω) = {(α(ω), β, γ(ω)) | α(ω)> x + β > y ≤ γ(ω) is valid for S(ω)}.
I The set S ] (ω) defines the feasible set we need in order to compute α(ω)
and γ(ω) for each ω ∈ Ω given β derived based on ω̄
I The following result guarantees the existence of such α(ω) and γ(ω)
Theorem
Let W be fixed for all ω ∈ Ω. Let x̂ ∈ X (binary) and ŷ ∈ YLP (ω̄, x̂) be given and
suppose that β > y ≤ g (ω̄, x̂, β) is valid for YIP
c
(ω̄, x̂). Then there exists a vector
α(ω̄) ∈ Rn1 and scalar γ(ω̄) ∈ R such that α(ω̄)> x + β > y ≤ γ(ω̄) is valid for
c
YIP (ω̄)
I The proof of this theorem provides a way for computing α(ω̄) and γ(ω̄),
restricted to some convex set Πx 0 .
21 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Given: (x̂, ŷ (ω̄)) ∈ X × YLP (ω, x̂) and β > y ≤ g (ω̄, x̂, β) that is valid for
c
YIP (ω̄, x̂)
Generate: α(ω̄)> x + β > y ≤ γ(ω̄) valid for YIP
c
(ω̄) by solving this LP
22 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
Given: (x̂, ŷ (ω̄)) ∈ X × YLP (ω, x̂) and β > y ≤ g (ω̄, x̂, β) that is valid for
c
YIP (ω̄, x̂)
Generate: α(ω̄)> x + β > y ≤ γ(ω̄) valid for YIP
c
(ω̄) by solving this LP
I Problem (7) characterizes the reverse polar S ] (ω̄) based on the disjunction
c
{−xi ≥ 0} ∪ {xi ≥ 1} and YIP (ω̄)
23 / 31
Introduction and Motivation
Closely Related Work Subgradient Decomposition
Methodology Fenchel Decomposition (FD)
Example Illustration Fenchel Disjunctive Decomposition (FDD)
Summary
FDD Algorithm
Step 0. Initialization.
Step 1. Solve Subproblem LPs.
I For each ω ∈ Ω solve subproblem (5). Step 3. Re-Solve Subproblem LPs
I If (x k , {y k (ω)}ω∈Ω ) satisfy integer I For each ω ∈ Ω solve the subproblem.
restrictions, compute optimality cuts for I Compute optimality cuts for E and D
E and D terms, update upper bound and
terms.
go to step 4. I If (x k , {y k (ω)}ω∈Ω ) satisfy integer
I Otherwise, go to step 2.
restrictions, update upper bound and go
Step 2. Generate Cutting Planes and Refine Feasible to step 4.
Region. I Otherwise, go to step 5.
I FD Cut: generate βk . Step 4. Termination.
I Form and solve separation problem to I Check termination criterion.
obtain βk and g (ω̄, αk (ω̄), βk ).
Step 5. Update and Solve the Master Problem.
I Form and solve reverse polar LP (7) to I Add optimality cuts to master problem
compute αk (ω) and γk (ω) for each (4).
ω ∈ Ω. I Solve master problem and update lower
I For each ω ∈ Ω, use βk from Step 2(i) bound.
to form problem (7) and solve to get I Repeat from Step 1.
αk (ω) and γk (ω).
I Form cut −βk> y ≥ −γk (ω) + αk (ω)> x k
and append to subproblem (5).
24 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Numerical Example
Instance Data:
c = 2; A = −1 ; b = −1
Problem:
60
q= ; p(ω1 ) = p(ω2 ) = 0.5;
20
Min E[f (x, ω̃)] + λD[f (x, ω̃)] 4
3
−1
s.t. Ax ≥ b − 5 −1
x ∈ {0, 1} W = −1
2
0
0 −1
where f (x, ω̃) = c > x + Q(x, ω̃) and
− 43 − 13
−2 −2
Q(x, ω) =Min q > y h(ω1 ) = −1 ; h(ω2 ) = −1
s.t. Wy ≥ h(ω) − T (ω)x −1 −1
y ∈ {0, 1}2
0 0
2 −1
D : Absolute semi-deviation (ASD) T (ω1 ) = 0 ; T (ω2 ) = 0
0 0
Consider λ = 0.5
25 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
DEP:
26 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
27 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Cut at k=1
28 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
k = 2: Updated feasible regions for scenarios 1 (ω1 ) and 2 (ω2 ) at x = 0
and x = 1
Cut at k=1
Cut at k=2
29 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Summary
Mean-risk SMIPs are difficult to solve and easily becoming intractable for
standard solvers.
In this talk, we present the Fenchel disjunctive decomposition (FDD)
method for SMIP with absolute semideviation which combines subgradient
optimization, Fenchel decomposition and disjunctive programming
The FDD cuts take advantage of the fixed recourse property and are
generated based only on a subset of scenarios via the common-cut
-coefficient approach.
Fenchel based cuts are generally expensive to compute are better suited for
problems with special structure such as fixed recourse involving knapsack
constraints and first- and second-stage binary decision variables
Practical applications and preliminary computational results motivated the
derivation of the FDD method – implementation is underway!
30 / 31
Introduction and Motivation
Closely Related Work
Methodology
Example Illustration
Summary
Thank You!
31 / 31