dropbox cec 2011 new orleans

Download Dropbox CEC 2011 New Orleans

If you can't read please download the document

Upload: maribel-garcia-arenas

Post on 26-May-2015

330 views

Category:

Technology


1 download

DESCRIPTION

How to use Dropbox (TM) for parallel evolutionary computation.

TRANSCRIPT

  • 1. Cloud-based EAs: An algorithmic study Juan-J. Merelo,Maribel Garca-Arenas, Antonio M. Mora,Pedro Castillo, Gustavo Romero,JLJ Laredo

2. IDEA

  • What do you know about cloud storage services?

3. Why not to use them for computing? 4. How can we use all our computers for make a multi-computer?

  • Desktop computer

5. Portable computer 6. Home computer 7. Any other ... 8. How to test the idea

  • Look for some storage FREE services
  • What are their _real_ services and what is the availability for
  • Store and Share information

9. Synchronization ? What about Dropbox? 10. DropboxTM

  • It is free up to a certain level of use (measured in traffic and usage)

11. It is popular, so many people use it, and we may found many volunteer for computation 12. It monitor the local filesystem and uploads information asynchronously 13. We can use it as a local directory 14. Putting in practice with Evolutionary Computation

  • What we need to make Evolutionary algorithms?
  • Exchange individuals: Phenotype and Genotype

We can exchange this information via files. So the name of the file represents the phenotype and genotype and all connected PCs can see it sharing with Dropbox 15. Lets go with the distributed algorithm

  • Dropbox makes the communications

16. It synchronizes the file-individuals with others computers 17. Each computer evolves an island 18. Dropbox folder is a pool of individuals and each computer adds and gets file-individuals from it 19. Lets go with the algorithm (II)

  • Each computer connected of synchronized by Dropbox is part of a multi-computer

20. All the computer evolves a population of individuals and it exchanges with the pool file-individuals 21. File-individuals

  • How to include phenotype and genotype into a file
  • Inside? It is not a good idea

22. Into the filesystem attributes? Dropbox is working on that and we will testing in the future 23. Into the namefile? It is our approach 24. File-individuals (II)

  • The namefile problems
  • How many gens can we include into the name?

25. We have to code the genotype into 32 base 26. Ex: 00000 -> 0, 00001-> 1, 01010->A ... 111111->V The name file includes: Fitness,genotypeBase32codification and the computer which generates the individual 27. Island Algorithm

  • Creates and evaluates the initial population

28. Until find the problem solution

  • Breed the population

29. Evaluate 30. Generational replacement with 1-elitism 31. If it is time, Gets and file-individual from the pool and incorporates it to the population 32. It it is time, Adds the best or a random file-individual to the pool Adds the best individual to the pool 33. Goals

  • What want we test?
  • It is Dropbox util for distributed evolutonary algorithms based on pool?

How can we test it?

  • Making an distributed evolutionary algorithm based on pool and demostrating that Dropbox synchronized architecture can run it as a multi-computer

34. Problems: MMDP

  • Multimodal Deceptive Problem

35. It is composed of k (k=80) subproblems of 6 bits each one called s i for i=0 to 79 . 36. Depending of the number of ones s itakes the values detailed into the table 37. The optimum fitness for this problem is 80 ones fitness 0 or 6 1 5 or 1 0 2 or 4 0,360384 3 0,640576 38. Problems: P-Peaks

  • This problem is created by generating P random N-bit strings where the fitness value of an individualis the number of bits that it has in common with the nearest peak divided by N

39. For this time, P is 300 and N is 600 and H is the hamming distance between x and the i-peak 40. The optimum fitness for this problem is 1. 41. Parameters

  • We use as multi-computer one, two or four heterogeneous computers so we use one, two or four island

42. Population size: 1000 individuals 43. Selection: 3-tournament 44. Crossover: uniform 45. Mutation: bit-flit 46. Replacement: Generational with 1-elitism 47. Stop criteria: maximum number of evaluations or to reach the solution of the problem 48. Parameters (II)

  • Migration
  • MMDP: This problem is solved usually around 3000 generations including only one computer and we will include in and out migration each 100, 200 and 400 generations so we migrate around 30, 15 or only 7 times during the evolution

49. P-Peaks:This problem is solved usually around 165 generations including only one computer and we will include in and out migration each 20, 40 and 60 generations so we migrate around 8, 4 or 2 times during the evolutionary process 50. First results: MMDP Computers Gens for Migration success(%) 1 100 83% 2 100 95% 4 100 100% 1 200 70% 2 200 88% 4 200 100% 1 400 80% 2 400 90% 4 400 100% 51. First results: P-PEAKS Computers Gens for Migration success(%) 1 20 100% 2 20 100% 4 20 100% 1 40 100% 2 40 100% 4 40 100% 1 60 100% 2 60 100% 4 60 100%

  • There is no difference for the success rate

52. Second results: MMDP 53. Second results: P-Peaks 54. Questions