Hello, There is lots of information about heavy (heavier) playouts on this mailing list. It looks like the main purpose of them is to make the evaluation better by killing dead groups with a high probability, connecting groups that are connected with high probability, etc. However, it looks like some randomness is desired, and mainly, "noise" should be reduced by pruning "bad" moves.
Which moves do you prune in the playouts? I saw some ideas including: -self atari -first line plays with no stones near -second/third line / near-center plays with no stones near -"bad" patterns (this seems vague to me) -ladders/escape sequences that don't work On the other hand, there are moves that should be played more often: -escape atari -kill groups -"good" patterns (vague again) -play in unsettled regions (for example determined by previous playouts) Are there others you would add? Also, it seems like a lot of experimentation is involved - people just try to find a good speed/strength balance. Also, how do you find "bad"/"good" patterns? Learn from pro games? Go knowledge? Secondly, how do you prune/favour these moves? I thought about 2 ideas: 1. Prune moves for the first x games. The higher x, the more certain we are that this move is worse than a random move. This way we might get better information initially but explore all moves eventually. 2. Prune moves with the probability p (0<=p<=1). This would lead to better playouts over all too, but the monte carlo tree is better balanced. Same goes for "good" moves, but instead of pruned, they are played deterministically. Additionally, I thought about checking the "good" moves in a different order every playout to add more randomness. This would be by creating permutations of the list of "features" (kill group, escape, etc.), then checking in that order. I would like to hear your opinions/ideas. ;) Cheers, ibd -- Psssst! Schon vom neuen GMX MultiMessenger gehört? Der kann`s mit allen: http://www.gmx.net/de/go/multimessenger _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
