I did play with the idea of different simple minded players for a version of MC that did not do a tree search and OGO for the Palm platform was based on this. The idea was AMAF with some number of global playouts but a priori biases were applied to moves depending on various factors such as whether they defended an atari, make a capture and other things. I did try a version that used a table like what is being described here.
I think the most powerful heuristic of all was not to play on the edge unless a stone was near. I don't remember whether "near" was a stone must not be touching or 1 point separation at most. One version used a table as described, containing weights for various points on the board to bias the choice but I don't remember that actually working, it did not make it to the final version. I don't think that is nearly so relevant if there is a stone nearby and I may not have tried that, I don't remember. This of course was not a playout rule, it was just a bias applied to all the root moves. All the biases together made a pretty big difference in the playing strength over simply picking the AMAF highest winning AMAF move that was legal. I assume the top programs using heavy playouts are doing a lot of that kind of stuff. Don On Thu, Dec 15, 2011 at 9:39 AM, Olivier Teytaud <[email protected]> wrote: > > > >> Strong players seem to adapt their tactics to the stones on the board. >> They are like a golfer with a diverse bag of clubs, who knows which one to >> swing. >> > > I've tried a UCB per player, choosing among a set of MC strategies. It > looks like your idea > of a golfer with a diverse bag of clubs. > Never succeeded. But maybe I've no imagination for building MC strategies. > Only Go players could design sequence-like patterns as Yizao did in the > first place > for MoGo (Yizao is 1st Dan or so, I think). > Olivier > > _______________________________________________ > 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
