> Part of the problem is that we're using global search to play a game > with lots of local little fights. If we have say five little fights, and > each takes say 11 ply to play well, it means a >50 ply global search > just to avoid local blunders. > > Currently, we try to sidestep this fundamental problem by replacing > local search with local knowledge, such as patterns. But that does not > fully use the power of search.
Hi Martin, I think Magnus is saying/doing the same thing, though it is stating the obvious to say it is hard to do right. Searching fights in isolation misses the interaction between the fights, which is one of the strengths of MCTS. I'd like to throw out an idea from one of my old programs, in case it can be adapted somehow. I attempted to learn patterns from pro games, and use that to search very deep (a bit like a MCTS playout, but I - foolishly - tried to do just one best playout). One of the main concepts was treating local sequences as a single move. A simple example is the 4-move hane-tsugi sequence [1] in the endgame. But it also applied to middle game invasion joseki, and opening joseki. So, I'd do all the tactical search, find the patterns (which were enriched with tactical chain status), then if a hane-tsugi pattern matched and was the most urgent pattern, play those 4 moves as if a single move. Then recalculate tactical search and repeat. This is obviously 4 times quicker, so allowed me to get deeper in the tree. (The patterns were fairly large, especially when combined with tactical information, so handle the more advanced [2] variations.) I also used these sequences to decide a just-right pattern size. So, we start with, say, a 3-manhatten distance around the first move. Then if the next move was within that distance treat it as a sequence, and merge in a 3-distance around that new stone as part of the pattern. When the pro game finally had a move outside the pattern area then the pattern is finished and stored. When using it, the combined pattern area must all match. No-sequence patterns were also added, and different patterns sizes can co-exist in the pattern database, so the tenuki moves also get considered (but usually with a lower pattern score). Darren [1]: http://senseis.xmp.net/?HaneTsugi [2]: http://senseis.xmp.net/?HaneTsugi%2FAdvanced -- Darren Cook, Software Researcher/Developer http://dcook.org/gobet/ (Shodan Go Bet - who will win?) http://dcook.org/mlsn/ (Multilingual open source semantic network) http://dcook.org/work/ (About me and my work) http://dcook.org/blogs.html (My blogs and articles) _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
