1 / 29

A Scalable Machine Learning Approach to Go

A Scalable Machine Learning Approach to Go. Pierre Baldi and Lin Wu UC Irvine. Contents. Introduction on Go Existing approaches Our approach Results Conclusion & Future work. What is Go?. What is Go?. Black & white play alternatively Stones with zero liberty will be removed

mloren
Télécharger la présentation

A Scalable Machine Learning Approach to Go

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Scalable Machine Learning Approach to Go Pierre Baldi and Lin Wu UC Irvine

  2. Contents • Introduction on Go • Existing approaches • Our approach • Results • Conclusion & Future work

  3. What is Go?

  4. What is Go? • Black & white play alternatively • Stones with zero liberty will be removed • The one who has more territory wins

  5. Why is Go interested? • Go is a hard game for computer. • The best Go computer programs are easily defeated by an average human amateur • Board games have expert-level programs • Chess: Deep blue (1997) & FRITZ (2002) • Checker: Chinook (1994) • Othello (Reversi): Logistello (2002) • Backgammon: TD-GAMMON (1992)

  6. Why is Go interested for AI? • Poses unique opportunities and challenges for AI and machine learning • Hard to build high quality evaluation function • Big branching factor, 200-300, compared with 35-40 for chess

  7. Existing approaches • Hard-coded programs • Evaluate the next move by playing large number of random games • Use machine learning to learn the evaluation functions

  8. Existing approaches── hard-coded programs • Hand-tailored pattern libraries • Hard-coded rules to choose among multiple hits • Tactical search (or reading) • E.g. “Many Faces of Go”, “GnuGo”

  9. Existing approaches── hard-coded programs • Pros: • Good performance • Cons: • Intensive manual work • Pattern library is not complete • Hard to manage and improve

  10. Existing approaches── Random games • Play huge number of random games from given position • Use the results of games to evaluate all the legal moves • Choose the legal move with best evaluation • E.g: Gobble, Go81

  11. Existing approaches── Random games • Pros • Easy to implement • Reasonable performance • Cons • Small boards only, cannot scale to normal board

  12. Existing approaches── Machine learning • Schraudolph et al., 1994 • TD0 • Neural Network • Graepel et al., 2001 • Condensed graph by common fate property • SVM • Stern, Graepel, and MacKay, 2005 • Conditional Markov random field

  13. Existing approaches── Machine learning • Pros: • Learn automatically • Cons: • Poor performance

  14. Out approach • Use scalable algorithms to learn high quality evaluation functions automatically • Imitate human evaluating process

  15. Our approach── Human evaluating process • Three key components • The understanding of patterns • The ability to combine patterns • The ability to relate strategic rewards to tactical ones

  16. Our approach── System components • 3x3 pattern library • Learn tactical patterns automatically • A structure-rich Recursive Neural Network • Propagate interaction between patterns • Learn the correlation between strategic rewards (Targets) and tactical reward (Inputs)

  17. Our approach── RNN architecture • Six planes • One input plane • One output plane • Four Hidden Planes

  18. Our approach── Update sequence

  19. Our approach── Provide relevant inputs • For intersections • Intersection type: black, white, or empty • Influence: influence from the same & opposite color • Pattern stability: a statistical value calculated from 3x3 patterns • For groups • Number of eyes • Number of 1st, 2nd, 3rd, and 4th order liberties • Number of liberties of the 1st and 2nd weakest opponents

  20. Our approach── Pattern stability (I) • 9x9 board is split into 10 unique locations for 3x3 patterns with mirror and rotation symmetries considered • Stability is measured for each intersection of each pattern within each unique location.

  21. Our approach── Pattern stability (II) • Ten unique pattern locations

  22. Our approach── Pattern stability (III)

  23. Our approach── Pattern stability results (I)

  24. Our approach── Pattern stability results (II)

  25. Results── Validation error

  26. Results── Results on move predictions

  27. Results── Matched move (I)

  28. Results── Matched move (II)

  29. Conclusion & Future work

More Related