A short description of Many Faces: It's an MCTS full board searcher. For the tree it uses the UCT formula and RAVE, with an exploration term, and an MFGO bias. It does progressive unpruning up to a maximum of 30 moves per position. The unpruning decision is based on rave and MFGO bias.
MFGO bias is based on the move values from the old Many Faces of Go move generator (a full board static evaluation (with local tactical search), big pattern databases, and expert system rules). The generator estimates points gained by each move. The playouts are pretty heavy, with local responses, hand tuned 3x3 patterns, and moves played with a probability distribution similar to Crazy Stones gamma values, but without the automatic learning. Gamma values are hand tuned. There are rules for not filling eyes, not making self Atari (unless it is a good self Atari), and avoiding other kinds of bad moves. The eye and self Atari rules are a little different from published methods. There is no tactical look-ahead in the playouts. The multithread search is different from other programs. It doesn't do pondering yet. David > -----Original Message----- > From: [email protected] [mailto:computer-go- > [email protected]] On Behalf Of Robert Jasiek > Sent: Wednesday, December 29, 2010 11:48 PM > To: [email protected] > Subject: Re: [Computer-go] News on Tromp-Cook ? > > Despite his loss of the bet on the surface, I congratulate Darren for > almost correctly predicting the 19x19 computer strength development! It > has been an extraordinarly impressive improvement during the last 3 > years! Before 19x19 was more like 10 kyu - now during parts of a game > ManyFaces can hold 1d to 2d level! With some more programming effort for > holding a program's playing strength at a constant level (maybe also by > filtering computer suggested moves by a human approach bias filter to > discard obviously bad moves like A15 in game 3 and by making endgame > more expert-orientated again), this strength can soon be held during an > entire game. > > Nick has said that a 2007 Hungarian RAVE paper was the theoretical > breakthrough. Is this its URL? > > http://zaphod.aml.sztaki.hu/papers/ecml06.pdf > > The site appears to be down though. Is there an alternative URL? > > ManyFaces was described as an expert system. How does it work today? How > does it use the modern algorithmic theories? > > Congratulations also to all the theorists! Without their great > discoveries, programs would still be weak. Might somebody please give an > overview on the relevant theories and how they work? > > One thing keeps bothering me though: What does all the strength > improvement give us humans for better understanding the game strategy? > Almost nothing? The information contained in the current calculation > size is not easily translated to human applicable strategic / tactical > knowledge. Other research, which is closer to the human way of go > understanding, by people like Berlekamp, Spight, Cazenawe or myself is > much more useful for players but its playing strength equivalent - > despite a few 10p knowledge exceptions - is still on the 20k level. > Currently there is an extreme gap between computer go theory making > computers strong, maths theory explaining go theory for human > understanding and traditional professional go theory, which fails to > explain well but allows eager and gifted players to succeed by means of > unlimited investment of time and effort. What is still mostly missing > are ways to link well to each other the three major paths towards great > playing strength. > > - How can programs learn well from professional knowledge? > - How can programs use well mathematical descriptions of human-like > strategy? > - How can players learn well from strong programs? > - How can further mathematical descriptions of human-like strategy be > derived from strong computer play or its underlying algorithms? > > Oh, and of course congratulations to John! > > -- > robert jasiek > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
