MOEAs AutomaticallyImprovingtheAnytimeBehaviour A .../file/REF_39.pdf ·...
Transcript of MOEAs AutomaticallyImprovingtheAnytimeBehaviour A .../file/REF_39.pdf ·...
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
1/30
Automatically Improving the Anytime Behaviourof Multiobjective Evolutionary Algorithms
Andreea Radulescu1 Manuel López-Ibáñez2Thomas Stützle2
1LINA, UMR CNRS 6241, Université de Nantes, Nantes, [email protected]
2IRIDIA, CoDE, Université Libre de Bruxelles (ULB), Brussels, [email protected],[email protected]
March 22nd, 2013
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
2/30 Summary
1 Introduction
2 Anytime optimization
3 Experimental setup
4 Results
5 Conclusion
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
3/30 IntroductionEvolutionary Algorithms
Population
Parents
Offspring
Initialization
TerminationSurvivor selection
Parent selection
Recombination
Mutation
Figure: General framework of Evolutionary Algorithms
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
4/30 IntroductionManual vs Automatic Tuning
Finding an appropriate parameter configuration for aspecific algorithm is a difficult optimization problem;
Many evolutionary algorithms are still manually tuned;
Parameter values are established by conventions, ad hocchoices and experimental comparisons on a limited scale;
Automatic tuning methods are available that configureevolutionary algorithms without much human effort.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
5/30IntroductionAutomatic TuningFinal quality
Figure: Quality of Pareto-front approximations through a run
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
6/30IntroductionAutomatic TuningAnytime behaviour
Figure: Anytime behaviour quality
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
7/30 IntroductionPurpose of the research
Parameter values greatly determine the success of an EAin reaching optimal solutions;
Improve the anytime behaviour of multi-objectiveevolutionary algorithms;
Find automatically the algorithm configurations thatproduce the best anytime behaviour.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
8/30 Summary
1 Introduction
2 Anytime optimization
3 Experimental setupTested MOEAsBenchmark problemsMonotonicity of hypervolume in MOEAThe irace package
4 ResultsTuning for anytime behaviourTuning for final quality
5 Conclusion
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
9/30 Anytime optimization
An anytime algorithm returns as high-quality solutions aspossible at any moment of its execution;
The evaluation of the anytime behaviour of an algorithmcan be seen as a bi-objective problem;
The non-dominated front so obtained reflects the qualityof the anytime behaviour of the algorithm.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
10/30 Anytime optimization
Figure: Anytime behaviour quality
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
11/30 Anytime optimization
Figure: Anytime behaviour quality
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
12/30 Anytime optimization
Figure: Anytime behaviour quality
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
13/30 Anytime optimization
Figure: Anytime behaviour quality
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
14/30 Summary
1 Introduction
2 Anytime optimization
3 Experimental setupTested MOEAsBenchmark problemsMonotonicity of hypervolume in MOEAThe irace package
4 ResultsTuning for anytime behaviourTuning for final quality
5 Conclusion
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
15/30 Tested MOEAs
IBEA (Indicator-Based Evolutionary Algorithm)it can be adapted to arbitrary preference information;does not require additional techniques in order to preservethe diversity of the population.
NSGAII (Nondominated Sorting Genetic Algorithm II)Uses a fast nondominated sorting procedure;Implements an elitist-preserving approach.
SPEA2 (Strength Pareto Evolutionary Algorithm 2)a fixed-sized archive method;a nearest neighbor density estimation technique.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
16/30 Tested MOEAsParameters
Name Role TypecrossoverProbability probability of mating two real
solutionsexternalMutationProbability probability of mutating real
a solutioninternalMutationProbability probability of mutating real
a variable in a solutionpopulationSize number of solutions integer
crossoverDistrIndex distance between the children integerand their parents
mutationDistrIndex distribution of the integermutated values
scalingFactor (IBEA) fitness scaling factor realarchiveSize (SPEA2) size of the archive integer
k(SPEA2) k-th nearest neighbour integer
Table: Parameters of IBEA, NSGAII and SPEA2
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
17/30 Benchmark problemsZDT
Scalable to any number of decision variables;Controlled difficulty in converging to the Pareto-optimalfront;Pareto-optimal front easy to construct.
ZDT1 convex ZDT2 nonconvex ZDT3 noncontiguous
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
18/30 Benchmark problemsDTLZ
+ scalable to any number of objectives.
DTLZ2 DTLZ5 DTLZ7 disconnected set
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
19/30 Monotonicity of hypervolume in MOEA
ZDT1 without an externalarchive
ZDT1 with an externalunbounded archive
DTLZ1 without an externalarchive
DTLZ1 with an externalunbounded archive
Figure: Monotonicity preservation of the hypervolume
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
ExperimentalsetupTested MOEAs
Benchmark problems
Monotonicity ofhypervolume inMOEA
The irace package
Results
Conclusion
20/30 The irace packageThe iterated racing algorithm
sampling: parameter configurations are sampled from aprobabilistic distribution (uniformly random at the start);
selecting: parameter configurations are run on a fewtraining problem instances, and statistically worseconfigurations are discarded until a few remain orcomputational budget exhausted;
updating: the sampling distribution is modified accordingto the selected configurations to bias sampling towardsbest configurations found.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
21/30 Summary
1 Introduction
2 Anytime optimization
3 Experimental setupTested MOEAsBenchmark problemsMonotonicity of hypervolume in MOEAThe irace package
4 ResultsTuning for anytime behaviourTuning for final quality
5 Conclusion
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
22/30 Tuning for anytime behaviourAnytime behaviour values
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●●
●●
●
●
● ●
●
●
●
●
●
●
0.6 0.7 0.8 0.9 1.0 1.1
0.6
0.7
0.8
0.9
1.0
1.1
AnytimeTuned
Def
ault
NSGAII DTLZ
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
● ●
● ●
●
● ●●
●
●
●
●
●●
●
●●
●
●
●
0.7 0.8 0.9 1.0 1.10.
70.
80.
91.
01.
1
AnytimeTuned
Def
ault
SPEA2 DTLZ
●
●
●●●
●
●
●
●
●●●
●●
●●●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
● ●
●●
0.8 0.9 1.0 1.1
0.8
0.9
1.0
1.1
AnytimeTuned
Def
ault
IBEA DTLZ
Figure: Anytime behaviour quality for the default configurationversus configurations tuned for anytime behaviour
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
23/30 Tuning for anytime behaviourQuality variation
0 500 1000 2000 30001.00
1.10
1.20
1.30
Evaluations
Hyp
ervo
lum
e
ParetoOptimalNSGAIIDefaultNSGAIIAnytimeNSGAIIFinal
0 500 1000 2000 30001.
051.
101.
151.
201.
251.
30
Evaluations
Hyp
ervo
lum
e
ParetoOptimalSPEA2DefaultSPEA2AnytimeSPEA2Final
0 500 1000 2000 3000
1.00
1.10
1.20
1.30
Evaluations
Hyp
ervo
lum
e
ParetoOptimalIBEADefaultIBEAAnytimeIBEAFinal
Figure: Variation of the quality of the Pareto front obtained for thethree different configurations
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
24/30 Tuning for anytime behaviourFinal values
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
1.0 1.1 1.2 1.3
1.0
1.1
1.2
1.3
AnytimeTuned
Def
ault
NSGAII DTLZ
●●
●
●
●
●
●
●●●●
●
●●●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
0.9 1.0 1.1 1.2 1.30.
91.
01.
11.
21.
3
AnytimeTuned
Def
ault
SPEA2 DTLZ
●●●●●●
●
●●●●●●●●●●●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●●●●
●
0.8 0.9 1.0 1.1 1.2 1.3
0.8
0.9
1.0
1.1
1.2
1.3
AnytimeTuned
Def
ault
IBEA DTLZ
Figure: Final Pareto front quality for the default configuration versusconfigurations tuned for anytime behaviour
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
25/30 Tuning for final qualityAnytime behaviour values
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
0.7 0.8 0.9 1.0 1.1
0.7
0.8
0.9
1.0
1.1
AnytimeTuned
Fin
alTu
ned
NSGAII ZDT
●
●
●
●●●
●
●
●
●●●
●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●
0.7 0.8 0.9 1.0 1.1
0.7
0.8
0.9
1.0
1.1
AnytimeTuned
Fin
alTu
ned
SPEA2 ZDT
●
●
●
●● ●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
0.7 0.8 0.9 1.0 1.1
0.7
0.8
0.9
1.0
1.1
AnytimeTuned
Fin
alTu
ned
IBEA ZDT
Figure: Difference for anytime behaviour quality
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
26/30 Tuning for final qualityFinal values
●●●
●●●
●
●
●●
●
●●
●●
●●
●
● ●●
● ●●
●
●
●
●
●
●
0.95 1.00 1.05 1.10 1.15 1.20
0.95
1.00
1.05
1.10
1.15
1.20
AnytimeTuned
Fin
alTu
ned
NSGAII ZDT
●●●
●
●●
●
●
●
●
●●●
●●
●
●●
●●
●●●●
●
●
●●●
●
0.6 0.7 0.8 0.9 1.0 1.1 1.2
0.6
0.7
0.8
0.9
1.0
1.1
1.2
AnytimeTuned
Fin
alTu
ned
SPEA2 ZDT
●●●●●●
●
●●
●
●
●
●
●●● ●●
●●
●●●●
●
●
●
●●●
0.7 0.8 0.9 1.0 1.1 1.2
0.7
0.8
0.9
1.0
1.1
1.2
AnytimeTuned
Fin
alTu
ned
IBEA ZDT
Figure: Difference for the quality of the final Pareto front
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
ResultsTuning for anytimebehaviour
Tuning for finalquality
Conclusion
27/30 Tuning for final qualityQuality variation
0 500 1000 2000 3000
1.00
1.10
1.20
1.30
Evaluations
Hyp
ervo
lum
e
ParetoOptimalIBEANSGAIISPEA2
Default conf.
0 500 1000 2000 30001.00
1.10
1.20
1.30
Evaluations
Hyp
ervo
lum
e
ParetoOptimalIBEANSGAIISPEA2
Anytime conf.
0 500 1000 2000 3000
1.15
1.20
1.25
1.30
Evaluations
Hyp
ervo
lum
e
ParetoOptimalIBEANSGAIISPEA2
Final conf.
Figure: Variation of the quality of the Pareto front approximationobtained by the three MOEAs
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
28/30 Conclusion
1 The quality of the anytime behaviour is significantlyimproved;
2 The tuned configurations improve the search behaviour ofMOEAs;
3 The MOEAs are more robust to specific terminationcriteria;
4 Substantial human effort is saved.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
29/30 Future work
−→ The impact of the anytime behaviour tuning can beextended:
Different categorical parameters;Analysis of others MOEAs;Tests with different benchmark problems.
AutomaticallyImproving the
AnytimeBehaviour of
MOEAs
A. Radulescu,M. López-Ibáñez, T.Stützle
Introduction
Anytimeoptimization
Experimentalsetup
Results
Conclusion
30/30
Automatically Improving the Anytime Behaviourof Multiobjective Evolutionary Algorithms
Andreea Radulescu1 Manuel López-Ibáñez2Thomas Stützle2
1LINA, UMR CNRS 6241, Université de Nantes, Nantes, [email protected]
2IRIDIA, CoDE, Université Libre de Bruxelles (ULB), Brussels, [email protected],[email protected]
March 22nd, 2013