Analysing the Performance of Different Population Structures for an Agent-based Evolutionary Algorithm
0 likes527 views
The document compares the performance of different population structures (ring, Watts-Strogatz, newscast) for a peer-to-peer evolutionary algorithm. It models a fine-grained parallel evolutionary algorithm using a peer-to-peer protocol as the underlying population structure. The experimental analysis tests these population structures on a 2-trap problem and finds that regular lattices require smaller population sizes but more evaluations to find a solution, while different small-world methods produce equivalent performance. Future work is proposed to validate the model on a real peer-to-peer infrastructure and explore other population structures and metaheuristics.
Analysing the Performance of Different Population Structures for an Agent-based Evolutionary Algorithm
- 1. Introduction P2P in a Nutshell The Evolvable Agent Experimental Analysis Analysing the Performance of Different Goals Methodology Analysis of Population Structures for an Agent-based Results Conclusions Evolutionary Algorithm Future Works Juan Luis Jim´nez Laredo et al. e Dpto. Arquitectura y Tecnolog´ de Computadores ıa Universidad de Granada 18-Jan-2011 1 / 17
- 2. Scope Introduction P2P in a Nutshell The Evolvable Agent Experimental Analysis Goals • Status: Peer-to-Peer Evolutionary Computation (P2P EC) Methodology Analysis of Results represents a parallel solution for hard problems Conclusions optimization Future Works • Modelling: Fine grained parallel EA using a P2P protocol as underlying population structure • Objective: Comparison of different population structures on the EA performance 2 / 17
- 3. Outline Introduction P2P in a Nutshell The Evolvable Agent 1 Introduction Experimental Analysis P2P in a Nutshell Goals Methodology The Evolvable Agent Analysis of Results Conclusions 2 Experimental Analysis Future Works Goals Methodology Analysis of Results 3 Conclusions 4 Future Works 3 / 17
- 4. Outline Introduction P2P in a Nutshell The Evolvable Agent 1 Introduction Experimental Analysis P2P in a Nutshell Goals Methodology The Evolvable Agent Analysis of Results Conclusions 2 Experimental Analysis Future Works Goals Methodology Analysis of Results 3 Conclusions 4 Future Works 4 / 17
- 5. P2P in a Nutshell Introduction P2P in a Nutshell The Evolvable Agent Experimental P2P EC Analysis Goals • Virtualization: Methodology Analysis of Results Single view at Conclusions application level Future Works • Decentralization: No central management • Massive Scalability: Up to thousands of computers 5 / 17
- 6. Outline Introduction P2P in a Nutshell The Evolvable Agent 1 Introduction Experimental Analysis P2P in a Nutshell Goals Methodology The Evolvable Agent Analysis of Results Conclusions 2 Experimental Analysis Future Works Goals Methodology Analysis of Results 3 Conclusions 4 Future Works 6 / 17
- 7. The Evolvable Agent Model Introduction P2P in a Nutshell Design principles The Evolvable Agent • Agent based approach Experimental Analysis • Fine grain parallelization Goals • Spatially structured EA Methodology Analysis of Results • Local selection Conclusions Future Works 7 / 17
- 8. The Evolvable Agent Model Introduction P2P in a Nutshell Design principles The Evolvable Agent • Agent based approach Experimental Analysis • Fine grain parallelization Goals • Spatially structured EA Methodology Analysis of Results • Local selection Conclusions Future Works 7 / 17
- 9. Outline Introduction P2P in a Nutshell The Evolvable Agent 1 Introduction Experimental Analysis P2P in a Nutshell Goals Methodology The Evolvable Agent Analysis of Results Conclusions 2 Experimental Analysis Future Works Goals Methodology Analysis of Results 3 Conclusions 4 Future Works 8 / 17
- 10. Goals and Test-Cases Introduction P2P in a Nutshell The Evolvable Agent Experimental Goal Analysis Goals • Comparison of performances using different population Methodology Analysis of Results structures Conclusions Future Works Ring Watts-Strogatz Newscast 9 / 17
- 11. Outline Introduction P2P in a Nutshell The Evolvable Agent 1 Introduction Experimental Analysis P2P in a Nutshell Goals Methodology The Evolvable Agent Analysis of Results Conclusions 2 Experimental Analysis Future Works Goals Methodology Analysis of Results 3 Conclusions 4 Future Works 10 / 17
- 12. Experimental settings Introduction P2P in a Nutshell The Evolvable Agent Experimental Analysis • 2-Trap. L=12...60 Goals Methodology • Population size Analysis of Results • Estimated by bisection Conclusions • Selectorecombinative Future Works GA (Mutation less) • Minimum population size able to reach 0.98 of SR • Uniform Crossover • Binary Tournament 11 / 17
- 13. Outline Introduction P2P in a Nutshell The Evolvable Agent 1 Introduction Experimental Analysis P2P in a Nutshell Goals Methodology The Evolvable Agent Analysis of Results Conclusions 2 Experimental Analysis Future Works Goals Methodology Analysis of Results 3 Conclusions 4 Future Works 12 / 17
- 14. Population Structure Introduction P2P in a Nutshell The Evolvable Agent Experimental Settings Analysis Goals Methodology Problem instance: 2-trap Analysis of Results Pop. Size: Tuning Algorithm Conclusions No Mutation Future Works 13 / 17
- 15. Population Structure Introduction P2P in a Nutshell Settings The Evolvable Agent Problem instance: L=60 2-trap Experimental Analysis Pop. Size: 135 Goals Methodology Max. Eval: 5535 Analysis of 1 Results Mutation: Bit-flip Pm = L Conclusions Future Works 14 / 17
- 16. Conclusions Introduction P2P in a Nutshell The Evolvable Agent Experimental Analysis Goals Methodology • Regular lattices require of smaller population sizes Analysis of Results ... BUT a bigger number of evaluations to find a solution. Conclusions • Different small-world methods produce an equivalent Future Works performance ...That’s good! Many P2P protocol are designed to work as small-world networks (i.e. Interoperability/Migration between P2P platforms) 15 / 17
- 17. Future Works Introduction P2P in a Nutshell The Evolvable Agent Experimental Analysis Goals Methodology Analysis of Results • Validation of the model in a real P2P infrastructure Conclusions • Exploration of other P2P protocols as population Future Works structures • Extension of the P2P concept to other metaheuristics 16 / 17
- 18. Questions Introduction P2P in a Nutshell The Evolvable Agent Experimental Analysis Goals Methodology Analysis of Results Conclusions Thanks for your attention! Future Works 17 / 17