Set-partitioning formulations for vehicle routing problems Strong degree constraints Computational experience Concluding remarks Strong degree constraints to impose partial elementarity in shortest path problems under resource constraints Claudio Contardo 1 , Guy Desaulniers 2 1 École des sciences de la gestion, UQÀM and GERAD 2 MAGI, École Polytechnique de Montréal and GERAD Column Generation 2012 Bromont, Canada, June 2012 Contardo and Desaulniers Strong degree constraints in SPPRC
36
Embed
Strong degree constraints to impose partial elementarity ... · Strong degree constraints to impose partial elementarity in shortest path problems under resource constraints Claudio
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Strong degree constraints to impose partialelementarity in shortest path problems under
resource constraints
Claudio Contardo1, Guy Desaulniers2
1École des sciences de la gestion, UQÀM and GERAD
2MAGI, École Polytechnique de Montréal and GERAD
Column Generation 2012Bromont, Canada, June 2012
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Outline
1 Set-partitioning formulations for vehicle routing problemsBasic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
2 Strong degree constraintsGeneral descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
3 Computational experience
4 Concluding remarks
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Some notation
V set of nodes in the graphVD set of depotsVC set of customers
R set of routes. Depending on how R is defined, routesmay contain cycles or notFor a given customer i ∈ VC and route r ∈ R we define
air : number of times that route r visits customer iξir : binary constant equal to 1 iff route r visits customer i
Basic Property
air ≥ ξir for all i ⊆ VC, r ∈ R (1)
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
r = (0,1,2,3,1,4,5,6,1,0)
a1r = 3
ξ1r = 1
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Example
0
1
2
3
4 5
6
r = (0,1,2,3,1,4,5,6,1,0)
a1r = 3
ξ1r = 1
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
General form of the set-partitioning formulation
min∑
r∈R
crθr + other terms (2)
subject to∑
r∈R
airθr = 1 i ∈ VC (3)
θr ∈ {0,1} r ∈ R (4)
other variables and constraints (5)
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
The Capacitated VRP (CVRP)
min∑
r∈R
cr θr (6)
subject to∑
r∈R
airθr = 1 i ∈ VC (7)
θr ∈ {0,1} r ∈ R (8)
Routes in set R are assumed to respect vehicle capacity Q
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
The VRP with Time Windows (VRPTW)
min∑
r∈R
cr θr (9)
subject to∑
r∈R
airθr = 1 i ∈ VC (10)
θr ∈ {0,1} r ∈ R (11)
Routes in set R are assumed to respect vehicle capacity Q andtime windows of customers [ei , li ]
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
The Multiple Depot VRP (MDVRP)
min∑
r∈R
cr θr (12)
subject to∑
r∈R
airθr = 1 i ∈ VC (13)
∑
r∈Ri
qrθr ≤ bi i ∈ VD (14)
θr ∈ {0,1} r ∈ R (15)
Routes in set R are assumed to respect vehicle capacity Q
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
The Capacitated Location-Routing Problem (CLRP)
min∑
i∈VD
fizi +∑
r∈R
crθr (16)
subject to∑
r∈R
airθr = 1 i ∈ VC (17)
∑
r∈Ri
qrθr ≤ bizi i ∈ VD (18)
θr ∈ {0, 1} r ∈ R (19)
zi ∈ {0, 1} i ∈ VD (20)
Routes in set R are assumed to respect vehicle capacity Q
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
And many others!
Inventory-Routing Problems
Periodic Routing Problems
Prize-collecting Routing Problems
And many combinations!
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Column generation for vehicle routing problems
Set-partitioning formulations provide stronger bounds thantraditional flow formulationsThe strength of set-partitioning formulations depends onthe set of feasible routes R
Pricing subproblem reduces to solve a shortest pathproblem with additional resource constraintsCurrent state-of-the-art methods find a balance betweenspeed and strength of the pricing subproblemBottleneck of column generaton methods is related to theimbalance between strength of the bounds vs. computingtimesMost efficient implementations use bidirectional dynamicprogramming in the pricing subproblem
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Basic concepts and notationGeneral formSeveral classes of vehicle routing problemsColumn generation for vehicle routing problems
Existing pricing algorithms
Without cycles elimination (SPPRC): very fast(pseudo-polynomial) but weak bounds
With 2-cycles elimination (2-cyc-SPPRC): slower butstronger bounds
With k-cycles elimination (k-cyc-SPPRC): slower butstronger bounds
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Strong degree constraints
It is a family of inequalities that are imposed into themaster problem
As such, they can be embedded into any columngeneration method
The goal is to impose partial elementarity in a smart way,taking advantage of master problem information (dualvariables)
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Strong degree constraints
Strong degree constraints (SDEG)∑
r∈R
ξirθr ≥ 1 for all i ∈ VC (21)
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Separation of SDEG
For each customer i ∈ VC we compute the quantity
v(i , θ) = 1 −∑
r∈R
ξirθr
We add to the RMP the SDEG constraints related to the kcustomers (k being a parameter defined a priori) with thelargest value of v(i , θ)
This is done if and ony if v(i , θ) ≥ ǫSDEG with ǫSDEG > 0being a parameter defined a priori
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Partial Elementarity
Let us consider a customer i ∈ VC. Let Rcyc(i) be the set ofroutes visiting customer i more than once. Assume that aSDEG is added for i .Let us subtract from the SDEG, the weak degree constraintassociated to customer i
∑
r∈R
(ξir − air )θr ≥ 0
⇒∑
r∈Rcyc(i)
(ξir − air )θr ≥ 0 (air = ξir for r ∈ R \ Rcyc(i))
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Partial Elementarity
Let us consider a customer i ∈ VC. Let Rcyc(i) be the set ofroutes visiting customer i more than once. Assume that aSDEG is added for i .Let us subtract from the SDEG, the weak degree constraintassociated to customer i
∑
r∈R
(ξir − air )θr ≥ 0
⇒∑
r∈Rcyc(i)
(ξir − air )θr ≥ 0 (air = ξir for r ∈ R \ Rcyc(i))
Theorem (Partial Elementarity)
θr = 0 for all r ∈ Rcyc(i) (air > ξir for r ∈ Rcyc(i))
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Effect of SDEG in the pricing subproblem
Let L, L′ be two labels (partial paths) sharing the sameterminal node v(L) = v(L′) = vLet PE ⊆ VC be the subset of customers for which aSDEG constraint has been addedLet PE(L),PE(L′) be the set of customers in PE served bythe partial path represented by L and L′, respectively
Regular Dominance Rule
L ≺ L′ ⇔ c(L) ≤ c(L′) and PE(L) ⊆ PE(L′) and o.c.
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
Effect of SDEG in the pricing subproblem
Let L, L′ be two labels (partial paths) sharing the sameterminal node v(L) = v(L′) = vLet PE ⊆ VC be the subset of customers for which aSDEG constraint has been addedLet PE(L),PE(L′) be the set of customers in PE served bythe partial path represented by L and L′, respectively
Regular Dominance Rule
L ≺ L′ ⇔ c(L) ≤ c(L′) and PE(L) ⊆ PE(L′) and o.c.
New (SHARPER!) Dominance Rule
L ≺ L′ ⇔ c(L) +∑
i∈PE(L)\PE(L′)
σi ≤ c(L′) and o.c.
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
General descriptionSeparation algorithmEffects of SDEG in the pricing subproblemStrong degree constraints vs. classic partial elementarity
We prefer to use SDEG instead of classic PEbecause...
1 dual variables of SDEG give us an estimate of theimportance of imposing PE on a certain node. Small dual⇒ elementarity is not that important ⇒ dominance rulebecomes much sharper
2 partial and total elementarity are particular cases in whichthe dual variables of SDEG constraints are "roughlyapproximated to infinity" in the dominance rule
3 in practice, solving a relaxation of the ESPPRC(2-cyc-SPPRC, ng-routes, k-cyc-SPPRC) suffices toobtain near-elementary routes after the addition of a fewSDEG
Contardo and Desaulniers Strong degree constraints in SPPRC
Set-partitioning formulations for vehicle routing problemsStrong degree constraintsComputational experience
Concluding remarks
Computational experience
We consider the VRPTW
The pricing subproblems are solved using bidirectionaldynamic programmingFive settings are tested