Pages

Sunday, March 13, 2016

Latest Predictions AlphaGo vs. Lee Sedol (9p)

               

[NL Versie] 




Nearly all polls worldwide show the same result:   ~75 - 85 % of all voters is convinced that Lee Sedol will win the coming match against AlphaGo. This is also the prediction of the majority of the participants of a price contest among dutch Go players (over  ~105 participants, organised in cooperation with chess and go shop  'het Paard ' and an ICT company). 




There are, however, many reasons to believe that AlphaGo, since the match about half a year ago, in which European Go Champion Fan Hui (2p) was beaten with 5 – 0, at least will play a few dan grades stronger against Lee Sedol. The estimated actual playing strength of AlphaGo is  ≥ 8th dan prof. 


  • AlphaGo has learned from it's (small) mistakes during the match against Fan Hui
  • improvements to AlphaGo's playing algorithms (for example for move selection and performance evaluation)
  • AlphaGo may be can built now on top of extended joseki and shape libraries 
  • finetuning and extension of AlphaGo's neural network training sessions
  • prevention and circumvention of specific problem situations (e.g. complex ko situations about many points)
  • selection of specific groups of professional Go games, not only from the KGS Go Server but as well (selectively) from other Go servers worldwide
  • improvement of the balance between on one hand AlphGo's neural networks for move selection and position evaluation and on the other hand precise computation through Monte Carlo Tree Search
  • extension of the number of conventional ( > 1202 CPUs) and graphical coprocessors (> 176 GPUs) that the distributed version of AlphaGo can use simultaneously during it's games against Lee Sedol
  • increase of the thinking time (computation time) per person: this match 2 hour p.p. (was 1 hour during the match against Fan Hui) which will be strongly beneficial for AlphaGo (especially towards the endgame)
  • implementation of new ideas and concepts to increase severely the performance of AlphaGo and/or make use of perhaps weaker elements in the way Lee Sedol plays (if these exist at all  since Lee Sedol has won over 68% of his games during ~the last years)
  • extension of the number of studied Go-positions (>60 million) and/or games played (e.g. against itself, ≥ 1.3 million) to increase the accuracy of AlphaGo in reproducing Go-profs moves. It has been shown by the DeepMind group that small improvements in this accuracy do lead immediately to big leaps forward in playing strength 
  • improvement and extension of position filters which determine whether a (subpart of a) position during a game against Lee Sedol is sufficiently being recognized by AlphaGo
  • improvements in reinforcement learning the value of Go moves by more detailed and accurate backpropagation of the final game result to each move and/or position


Despite all possible improvements of AlphaGo during the match against Lee Sedol, I expect that: 

  • Lee Sedol wins at least one game against AlphaGo 
  • the winner of the first game will also be the final winner of the match
  • Lee Sedol will win the match with 3-2 
  • AlphaGo also will win at least one game 
  • Lee Sedol demonstrable (*) will forget to play at least one move that he really had to play immediately, and this will happen at least once EACH game of the match
  • in the endgame, Lee Sedol demonstrable (*) plays less well than AlphaGo and, consequently, in EACH game of the match Lee Sedol will loose points in the endgame
  • at least one game Lee Sedol will reach a position of total resignation (*). Whether he indeed resigns or ultimately wins / loses is irrelevant
  • Lee Sedol has to put in all effort to realize and keep his eventually built profit until the end of the game
  • within one year after the match with Lee Sedol, AlphaGo will play a similar match against a strong 9p prof (possibly Lee Sedol again) and this time AlphaGo will win this match (of the five formal games, AlphaGo will win then at least 3 games).

(*) as for example reviewed by a majority of 10 independent top Go profs (9p) from South-Korea, China and Japan.






And if you do want to make bets about this:  Deep Learning models are as good as the data you feed them. 

No comments:

Post a Comment