> > > > I think there is a semi-successful program that uses (or did use) > traditional alpha/beta pruning. Was it Aya? > > I played with the idea of using playouts as the evaluation function of a > traditional chess-like search. I did not develop the idea but I think > there could be merit. One key insight here is that you do not need to > always do a lot of playouts. In a traditional PVS style chess search you > are trying to prove a node is above or below beta and that's all, so the > playouts can stop as soon as you have some confidence one way or the other. > > > Another insight is that modern chess programs have a pretty narrow tree, > the branching factor of my very strong program Komodo is less than 2. > You can still have a pretty low branching factor go program. I'm > pretty much convinced that MCTS and modern chess programs have less > differences that we imagine. I have not proved this but I believe the > pruning can be more severe in GO when it comes to techniques like late move > reductions, but on the other hand a very powerful technique in chess is > null move pruning which I don't think will work well in go. There is > a substitute that might however which is called "multi-cut" which is very > nearly as effective in chess and might work in go. > > Don >
Hi Don! I pretty much agree what you say. I think that a common misconception is to think that because Go has much greater branching factor Chess, it's more difficult for computers. Bullshit, there are very powerful methods to reduce effective branching factor which could be easily applied to Go. The only thing we are missing is a sane evaluation function. I'm a very weak player, but even then I am able to understand that it's almost hopeless tasks to measure algorithmically concepts like form, thickness, aji, effectiveness etc. Cheers, Joona
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
