Automatic Design of Ant Algorithms with Grammatical Evolution

Talk given at EuroGP in Málaga, Spain, April 13, 2012. The paper received the "Best Paper Award".

Transcript

1. Automatic Design of Ant Algorithms with Grammatical Evolution Jorge Tavares Francisco B. Pereira CISUC, University of Coimbra, Portugal
2. how do we design an Ant algorithm?
3. “A man provided with paper, pencil, and rubber, and subject to strict discipline, is in effect a universal machine.” Alan Turing
4. how can we automatically design an Ant algorithm?
5. Grammatical Evolution + Ant Structures = Evolutionary Ant Algorithms
6. running the system Initial Population Access Fitness of Population Satisfactory Individual exist ? Select Individuals Make Random Changes Return Best Individual Ant Algorithms Run Ant Algorithm to solve a problem Best Ant Algorithm for the problem
7. <aco> ::= (aco <parameters-init> <optimization-cycle>) <optimization-cycle> ::= (repeat-until <loop-ants> <update-trails> <daemon-actions>) <loop-ants> ::= (foreach-ant make-solution-with <decision-policy> (if <bool> (local-update-trails <decay>))) <decision-policy> ::= (roulette-selection) | (q-selection <q-value>) | (random-selection) <update-trails> ::= (progn <evaporate> <reinforce>) ... ...
8. the system is able to generate all main Ant algorithms Ant System (AS), Elitist Ant System (EAS), Rank- Based Ant System (RAS), Ant Colony System (ACS), Max-Min Ant System (MMAS)
9. the search space contains many other combinations that define alternative Ant algorithms
10. does search converge to manually designed Ant algorithms?
11. can we find novel Ant algorithms?
12. test problem: traveling salesman problem selected instances from TSPLIB: learning: eil76, pr76, gr96 testing: att48, eil51, berlin52, kroA100, lin105, gr137, u159, d198, pr226, lin318
13. 0! 0.05! 0.1! 0.15! 0.2! 0.25! 0.3! 0.35! 1! 5! 9! 13! 17! 21! 25! Distance to optimum! Generations! median of best individuals
14. evolution summary (30 runs)
15. search does not converge to manually designed architectures
16. main components frequency
17. evolved example: ei7618
18. but... are the evolved Ant algorithms effective?
19. optimization results
20. statistical analysis
21. comparison with standard algorithms
22. statistical analysis
23. hybrid between the two best evolved
24. 1) evolution is able to discover original architectures 2) best evolved strategies exhibit a good generalization capability 3) competitive with human-designed variants 4) hybrid architecture based on evolved strategies is effective conclusions
25. “We can only see a short distance ahead, but we can see plenty there that needs to be done.” Alan Turing
26. Thank you. Questions? Special thanks to Alex Wild for all the Ant pictures @alexanderwild.com