Hi, At least these two papers can be classified as "killer heuristics".
The first one is based on the rave values : http://hal.inria.fr/index.php?halsid=hlhgjavhdfnf5iren0tvsq4lg2&view_this_doc=inria-00485555&version=1 Here, the value of the "rave heuristic" is also used to bias the future decisions taken in the next Monte-Carlo simulations. The second one can be found here : http://hal.inria.fr/index.php?halsid=hlhgjavhdfnf5iren0tvsq4lg2&view_this_doc=inria-00456422&version=2 Here, for each node n and for each pair of moves (a,b) we keep in memory the number of won games and the number of losses where a and b have been played after n. Then, we compute the success rate when a and b have been played by the same player after n. Finally, we use this score to bias the next Monte-Carlo simulations. Fabien 2011/5/25 Stefan Kaitschick <[email protected]> > > In this game, there was a big semeai on the left side. The result was a >>> won >>> position for Aya, but both Aya and ManyFaces thought that ManyFaces had >>> won >>> (or perhaps that it was a semeai), so eventually Aya resigned before it >>> was >>> played out. >>> >> >> This position reminds me a similar situation in the game between Zen and >> Mogo in last TAAI tournament, >> >> http://files.gokgs.com/games/2010/11/18/mogobot5-DeepZen19.sgf >> >> It was Zen's big win, but it lost by misreading the "simple" semeai at >> the top-left corner. At move 351, Zen played D19 capture and the semeai >> became seki. But if Zen played D17 connect instead, it was ahead by around >> 50 points. >> >> Aja >> >> > That's a great example. The position is extremely simple, except for the > semeai. > The losing move creates liberties, while the winning move seems to do > nothing. > In this case there must have been specific rules in place to outlaw the > correct move. > Even a very shallow search would have quickly confirmed success. > > As to the problem of smarter playouts, has something along the lines of the > "killer heuristic"(the most successful response sofar to a specific move) , > used in chess programming, been tried? > There's a lot more housekeeping than with RAVE and AMAF, but I can't > believe it hasn't been tried. > Maybe the effort could be reduced by only storing the response-successrates > of the 5*5 surrounding area, or something like that. > Anyways, I haven't even heard of a failed attempt, which is a little > strange to me. > > Stefan > > _______________________________________________ > 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
