influence of the population structure on the performance of an agent-based evolutionary algorithm
TRANSCRIPT
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Influence of the Population Structure on thePerformance of an Agent-based Evolutionary
Algorithm
J.L.J. Laredo et al.
Dpto. Arquitectura y Tecnologıa de ComputadoresUniversidad de Granada
11-Sept-2010
1 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Scope
• Status: Peer-to-Peer Evolutionary Computation (P2P EC)represents a parallel solution for hard problemsoptimization
• Modelling: Fine grained parallel EA using a P2P protocolas underlying population structure
• Objective: Comparison of different population structureson the EA performance
2 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Outline
1 Introduction
2 Model DesignThe Evolvable Agent
3 Experimental AnalysisGoalsMethodologyAnalysis of Results
4 ConclusionsConclusions
5 Future Works
3 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Introduction
P2P EC
• Virtualization:Single view atapplication level
• Decentralization:No centralmanagement
• Massive Scalability:Up to thousands ofcomputers
4 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Population Structure as a complex network
Panmictic Small-world Regular lattice
n(n−1)2
log(n) n
5 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Population Structure as a complex network
Panmictic Small-world Regular lattice
n(n−1)2
log(n) n
5 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Population Structure as a complex network
Panmictic Small-world Regular lattice
n(n−1)2
log(n) n
5 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Population Structure as a complex network
Panmictic Small-world Regular lattice
n(n−1)2
log(n) n
5 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Outline
1 Introduction
2 Model DesignThe Evolvable Agent
3 Experimental AnalysisGoalsMethodologyAnalysis of Results
4 ConclusionsConclusions
5 Future Works
6 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
The Evolvable Agent Model
Design principles• Agent based approach
• Fine grain parallelization
• Spatially structured EA
• Local selection
7 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
The Evolvable Agent Model
Design principles• Agent based approach
• Fine grain parallelization
• Spatially structured EA
• Local selection
7 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Outline
1 Introduction
2 Model DesignThe Evolvable Agent
3 Experimental AnalysisGoalsMethodologyAnalysis of Results
4 ConclusionsConclusions
5 Future Works
8 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Goals and Test-Cases
Goal
• Comparison of performances using different populationstructures
Ring Watts-Strogatz Newscast
9 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Outline
1 Introduction
2 Model DesignThe Evolvable Agent
3 Experimental AnalysisGoalsMethodologyAnalysis of Results
4 ConclusionsConclusions
5 Future Works
10 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Experimental settings
• 2-Trap. L=12...60
• Population size• Estimated by bisection• Selectorecombinative
GA (Mutation less)• Minimum population
size able to reach 0.98of SR
• Uniform Crossover
• Binary Tournament
11 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Outline
1 Introduction
2 Model DesignThe Evolvable Agent
3 Experimental AnalysisGoalsMethodologyAnalysis of Results
4 ConclusionsConclusions
5 Future Works
12 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Population Structure
Settings
Problem instance: 2-trapPop. Size: Tuning AlgorithmNo Mutation
13 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Population Structure
Settings
Problem instance: L=60 2-trapPop. Size: 135Max. Eval: 5535Mutation: Bit-flip Pm = 1
L
14 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Conclusions
• Regular lattices require of smaller population sizes... BUT a bigger number of evaluations to find a solution.
• Different small-world methods produce an equivalentperformance...That’s good! Many P2P protocol are designed to workas small-world networks(i.e. Interoperability/Migration between P2P platforms)
15 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Future Works
• Validation of the model in a real P2P infrastructure
• Exploration of other P2P protocols as populationstructures
• Extension of the P2P concept to other metaheuristics
16 / 18
Introduction
Model Design
The EvolvableAgent
ExperimentalAnalysis
Goals
Methodology
Analysis ofResults
Conclusions
Conclusions
Future Works
Questions
Thanks for your attention!
17 / 18