On Wed, Apr 8, 2009 at 10:46 AM, Darren Cook <[email protected]> wrote: >> End game is another issue. MC programs only aim on winning, so they >> endgame is nor perfect in sense human would define it, but perfect >> enough to win if the game is winnable. > > You can modify komi to get the human expert and MC program in agreement. > > This suggests how you could automate a set of endgame problems: run Mogo > (or similar) with lots of time and keep increasing/reducing the komi > until you get a sudden swing in winning probability (e.g. from 60% to > 20% for side to play). That should tell you the komi to use, and at > least one of the winning moves. > > To find alternative winning moves you'd need to have some way to tell > Mogo, or whatever, it cannot choose the existing winning move you have, > and must choose a different move. Once winning probability suddenly > drops again it tells you there are no more winning moves left. > > If a program can generate the test set, why bother? I think it is useful > because you can tune against it. E.g. if you give Mogo 120s per move to > generate the test suite moves, then you tune until it can find the > correct moves for the whole test suite on 1s per move. > > Tuning the endgame play is very important for MCTS search, because every > playout always goes to the endgame. Strong endgame play in the playouts > should make a program stronger at all stages of a game. > > What do you think? Is such a endgame problem suite useful?
What I did to generate a test suite was to make Mogo play against itself (on 9x9, 10s per move) and stop the game when it thinks was clearly won for one of the player. That created a few hundred end game positions where the result were clear enough, and I used that test suite to evaluate the playouts. It was not precise enough (or I did not have enough positions?) to actual optimize details of the playouts, but it could make the difference between a clearly better/worse policy. However, I don't think it would be suitable to test the strength of the actual player, but only the playout policy. Sylvain > Darren > > -- > Darren Cook, Software Researcher/Developer > http://dcook.org/mlsn/ (English-Japanese-German-Chinese-Arabic > open source dictionary/semantic network) > http://dcook.org/work/ (About me and my work) > http://dcook.org/blogs.html (My blogs and articles) > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
