Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Parallel Problem Solving from Nature session-4


Published on

Introduction to 14 papers in session 4 of PPSN

Published in: Education
  • Be the first to comment

  • Be the first to like this

Parallel Problem Solving from Nature session-4

  1. 1. Session 4
  2. 2. View from above
  3. 3. We are all connected
  4. 4. Let's solve the Travelling Posterviewer problem
  5. 5. Careful! Don't touch here!
  6. 6. Paper #190: Evolutionary Detection of New Classes of Equilibria. Application in Behavioral Games D. Dumitrescu et al. <ul><li>People play games in a weird way.
  7. 7. A model has to be found to explain that behavior.
  8. 8. Multiobjective algorithm is used to evolve strategies that approach that behavior </li><ul><li>Which MOEA? </li></ul><li>For those into games </li></ul>
  9. 9. Paper #68: Evolving a Single Scalable Controller for an Octopus Arm with a Variable Number of Segments Woolley & Stanley. <ul><li>Size is not so important </li><ul><li>In terms of #dimensions of state/action space </li></ul><li>A indirectly encoded neurocontroller (HyperNEAT) for an octopus arm is evolved </li><ul><li>Fitness: ability in grasping a target </li></ul><li>A scalable neurocontroller architecture is found.
  10. 10. Next: application to Paul </li></ul>
  11. 11. Paper #104: Evolving Strategies for Updating Pheromone Trails: a Case Study with the TSP Tavares & Pereira <ul><li>Let GP devise your ACO pheromone updating strategy for you.
  12. 12. Weird and interesting update strategies arise </li><ul><li>They generalize better than other update strategies </li></ul><li>Terminal set is key (but not too open to experiments) </li></ul>
  13. 13. Paper #126: Globally Induced Model Trees: An Evolutionary Approach Czajkowski & Krętowski <ul><li>Regression trees useful for extracting patterns in databases </li><ul><li>What are regression trees? </li></ul><li>Usual: top-down approach. Bad. Here: global approach based on EA </li><ul><li>Heuristic mutation/crossover operators used
  14. 14. Akaike information criteria used as fitness </li></ul></ul>
  15. 15. Paper #211: Scheduling English Football Fixtures over the Holiday Period Using Hyper-heuristics Gibbs et al. <ul><li>Problem: producing a fixture schedule for English football </li><ul><li>Large problem space! </li></ul><li>Solucion: selection hyperheuristics </li><ul><li>Applying heuristics over single candidate solution </li></ul><li>Reinforcement learning rocks! </li></ul>
  16. 16. Paper #162: Evolutionary Learning of Technical Trading Rules without Data-mining Bias. Agapitos et al. <ul><li>Get rich quick! </li><ul><li>Predict performance of trading rules from past behavior. </li><ul><li>Rule performance decays over time </li></ul></ul><li>Grammar-based GP used for rule induction.
  17. 17. MOEA using returns + statistical significance of rule.
  18. 18. Encouraring results! </li></ul>
  19. 19. Paper #208: Differential Evolution Algorithms with Cellular Populations Dorronsoro & Bouvry <ul><li>DE based on updating real-number vectors via combinations of others </li><ul><li>Rarely used in a distributed environment. </li></ul><li>Several DE algorithms tested on a cellular framework.
  20. 20. Cellular better than not (except for JADE) </li></ul>
  21. 21. Paper #70: Hybrid Directional-biased Evolutionary Algorithm for Multi-objective Optimization Shimada et al. <ul><li>Nobody searches the edges of the Pareto front! </li><ul><li>Pareto would be upset with this </li></ul><li>hIDEA adds weights to objectives to allow this. </li><ul><li>Based on IBEA </li></ul><li>(Inverse) General Distance used to evaluate solutions
  22. 22. hIDEA searches wider and closer to Pareto front </li><ul><li>Pareto would be happy! </li></ul></ul>
  23. 23. Paper #80: Solving Multiobjective Optimization Problem by Constraint Optimization Jiang & Zhang & Ren <ul><li>Multiobjective turned into constraint optimization problems </li><ul><li>Multiobjective EA based on constraint optimization: MEACO
  24. 24. n objectives -> n – 1 constraints + 1 objective
  25. 25. COP solution not shown in paper
  26. 26. 2 objectives only? </li></ul><li>Spacing/convergence metrics measured. </li><ul><li>MEACO is the best almost always </li></ul></ul>
  27. 27. Paper #228: Environment-driven Embodied Evolution in a Population of Autonomous Agents Bredeche & Montanier <ul><li>Implicit fitness
  28. 28. Selfish gene hypothesis: strategy successful if it has been able to spread over a population.
  29. 29. Autonomous agents broadcast their genome while alive, load variation of one of those received. </li><ul><li>Implicit selection </li></ul><li>Autonomous agents improve! </li></ul>
  30. 30. Increasing exploration
  31. 31. Paper #131: One-Point Geometric Crossover Alberto Moraglio <ul><li>Crossover can be interpreted geometrically </li><ul><li>And generalized to any data structure where d is defined. </li></ul><li>Crossover results are actually placed over a geodesic between parents.
  32. 32. Crossover operators are proposed for diferent data structures </li><ul><li>And proofs made in flyspeck font </li></ul></ul>
  33. 33. Paper #146: Mirrored Sampling and Sequential Selection for Evolution Strategies Brockhoff & al. <ul><li>Did you know that single-parent non-elitist evolution strategy (ES) can be locally faster than the (1+1)-ES? </li><ul><li>But only with mirrored sampling and sequential selection! </li></ul><li>Convergence speed theoretically computed.
  34. 34. Sequential, mirrored versions of ES and CMA-ES faster and more robust </li></ul>
  35. 35. Paper #120: More Effective Crossover Operators for the All-Pairs Shortest Path Problem Doerr & al. <ul><li>APSP consists in computing all shortests paths betwen pair of nodes in a weighted graph.
  36. 36. Crossover with repair mechanisms improves optimization time.
  37. 37. Feasible parent solution even more: O n 3 log n </li></ul>
  38. 38. Paper #23: General Scheme for Analyzing Running Times of Parallel Evolutionary Algorithms Lässig & Sudholt <ul><li>Theoretical analysis of speed-up of parallel evolutionary algorithms on several functions and over different topologies.
  39. 39. Method based on fitness-level methods. </li><ul><li>Finding lower-bound for probability of leaving one level </li></ul><li>Differences abound. </li></ul>
  40. 40. Is source code available? No, it's not! That's not good! Except papers #228 and #146!
  41. 41. Thank you very much Happy postering!