Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Comparison of Crossover Operators
for the Capacitated Vehicle Routing Problem
Marek Kubiak
Faculty of Computer Science and Management
Poznań University of Technology
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Outline
1
Systematic construction of recombination operators for problems
of combinatorial optimisation
2
The Capacitated Vehicle Routing Problem (CVRP)
3
The crossover operators
4
Comparison experiments
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of operators: motivation
metaheuristic algorithms are not general tools of optimisation
(the No Free Lunch Theorem)
the way a given metaheuristic is adapted to a problem
may have crucial influence on the algorithm’s performance
domain knowledge has to be introduced in algorithms
for the sake of efficiency
schemes of systematic and justified adaptation
of metaheuristic algorithms to problems are welcome
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of operators: the procedure
make hypotheses about features of solutions which may
influence the goal function
define/implement distance functions which exploit the features
test the distance functions against global convexity
global convexity is revealed ⇒ implement recombination
operators which preserve the distances (DPX)
Global convexity examination is made
a tool of searching for important features of solutions
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
The Capacitated Vehicle Routing Problem (CVRP)
Input:
one depot
set of geographically
distributed customers with
certain demands
travel cost between
customers and the depot
set of identical vehicles with
limited capacity
Goal: find a set of routes which
satisfies customers’ demands and
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of operators for the CVRP
two features of solutions examined:
clusters of customers: sim
cpn
(s
1
, s
2
)
edges: sim
ce
(s
1
, s
2
)
high absolute values of correlation in sets of 1000 local optima
for most of instances found
3 DPX operators implemented:
CPX: preserves clusters of customers
CEPX: preserves common edges
CECPX: preserves both
a mutation operator CPM implemented which preserves
clusters of customers
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of operators for the CVRP
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of operators for the CVRP
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Systematic construction of operators for the CVRP
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
Crossover operators for the CVRP
route-based crossover (
RBX
), sequence-based crossover (SBX)
Potvin J.-Y., Bengio S., The Vehicle Routing Problem with Time Windows - Part II: Genetic Search,INFORMS Journal of Computing, 8, 2, 1996.
Jozefowiez N. et al., Parallel and Hybrid Models for Multi-Objective Optimization: Application to the Vehicle Routing Problem, in: Parallel Problem Solving from Nature - PPSN VII, Springer-Verlag, 2002
generic crossover (GCX), specific crossover (SCX)
Tavares J. et al., Crossover and Diversity: a Study about GVR, GECCO 2003.
split crossover (
SplitX
)
Prins C., A Simple and Effective Evolutionary Algorithm for the Vehicle Routing Problem, 4th
Metaheuristics International Conference, July 2001.
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
RBX: route-based crossover
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
SplitX: Prins’ representation
a permutation only
optimal division into
routes
arbitrary merge points
makes no difference for
local optima
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
SplitX: split crossover
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
SplitX: split crossover
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
CPX2: clusters preserving crossover
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
RBX SplitX CPX2 CEPX CECPX2
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed Preservation of distances Optimisation efficiency
Aspects of comparison
evolvability
speed
preservation of distances
optimisation efficiency
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Motivation for measuring evolvability
is an operator able to introduce something new to a
population?
what is the probability (for the given operator) to produce an
offspring better than its parents?
Altenberg L., The Schema Theorem and Price’s Theorem. In: Whitley D., Vose M. (eds.), Foundations of
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
The evolvability experiment
1000 trials of each crossover on the same set of local optima
after each crossover perform local search with 3 merged
neighbourhoods:
merge 2 routes
exchange 2 edges
swap 2 customers
examine the offspring after local search:
is it different from parents?
is it better than the worse parent?
(an approximation of an acceptance criterion)
is it better than the better parent?
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Evolvability: results
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Evolvability: results
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Evolvability: results
Op1>Op2
RBX
SplitX
CPX2
CEPX
CECPX2
Sum
RBX
0
4
-3
-20
-19
-38
SplitX
-4
0
-5
-20
-19
-48
CPX2
3
5
0
-21
-21
-34
CEPX
20
20
21
0
10
71
CECPX2
19
19
21
-10
0
49
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
The speed experiment
1000 trials of each crossover on the same set of local optima
after each crossover perform local search with 3 merged
neighbourhoods:
merge 2 routes
exchange 2 edges
swap 2 customers
examine the number of iterations of local search
compute the average time of one crossover (
to be done
)
investigate the relationship between the number of iterations
and quality of the offspring (
to be done
)
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Speed: results
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Motivation for measuring the preservation of distances
preservation of certain features in a VRP
Jaszkiewicz A., Kominek P., Genetic local search with distance preserving recombination operator for a vehicle routing problem, European Journal of Operational Research, 151, 2003, 352-364
preservation of certain features in the ROADEF 2003
Challenge problem
Jaszkiewicz A., Adaptation of the genetic local search algorithm to the management of Earth observation satellites, in: 7th National Conference on Evolutionary Computation and Global Optimization, Kazimierz
Dolny, Poland, 2004
important features for the CVRP were identified
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
A distance metric for sets
A normalized distance metric:
dists
(S
1
, S
2
) =
|S
1
∪ S
2
| − |S
1
∩ S
2
|
|S
1
∪ S
2
|
=
=
|S
1
| + |S
2
| − 2 · |S
1
∩ S
2
|
|S
1
| + |S
2
| − |S
1
∩ S
2
|
=
=
18 + 18 − 2 · 9
18 + 18 − 9
=
2
3
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
A distance metric for clusterings
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
A distance metric for clusterings
distc
(D
1
, D
2
)
=
1/2 · max
C
1i∈D
1{min
C
2j∈D
2{dist
s
(C
1i
, C
2j
)}} +
+1/2 · max
C
2j∈D
2{min
C
1i∈D
1{dist
s
(C
1i
, C
2j
)}}
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
A distance metric for clusterings
dist
s
C
11
C
12
C
13
min
C
21
0.29
1
1
0.29
C
22
0.86
0.93
0.69
0.69
C
23
1
0.44
0.85
0.44
min
0.29
0.44
0.69
distc
=0.69
Table:
Computation of dist
c
for the exemplary clusterings.
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
The distance preservation experiment
1000 trials of each crossover on the same set of local optima
after each crossover perform local search with 3 merged
neighbourhoods:
merge 2 routes
exchange 2 edges
swap 2 customers
examine the distances:
between parents
between the offspring and its parents
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Distance preservation: initial results
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
Distance preservation: initial results
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Evolvability Speed
Preservation of distances Optimisation efficiency
The optimisation efficiency experiment
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Conclusions
high evolution potential in systematically constructed
operators
the fastest operators are not necessarily the best
forbidding LS moves should be beneficial (preserve distance,
speed up computation)
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research
Further research
process results of distance preservation experiments
perpare optimisation efficiency experiment
implement and compare to other operators
evolvability:
how does the probability change with increasing quality of
solutions?
does the probability correlate with strength of global convexity?
implement the forbidding of moves in local search
exploit distance metrices in convergence analysis of the
memetic algorithm
Systematic construction of recombination operators. . . The Capacitated Vehicle Routing Problem (CVRP) The crossover operators Comparison experiments Conclusions and further research