Is UCT really that good at finding the best move in alot of situations? A friend of mine pointed out that in alot of situations it would seem that there would be alot of statistical "noise" in the system. Purely random playouts from a starting position even on 9x9 may not yield alot of information using a pure UCT bot(which may be why people are resorting to opening books to improve performance). Alot of the outcomes of the playouts might be purely random or dependent on mistakes made far down the line from the move under analysis.
One thing I saw mentioned in Sylvain's thesis (chapter 4) was that reasonable sequence like behavior seems to greatly improve the strength of the bot so you have to craft the pattern libraries to get this sort of behavior and combine it with a proximity style heuristic. The published materials also suggests he uses a dragon heuristic and of course inherits the atari/capture heuristic from first generation Mogo. Obviously testing is probably the only answer and this points to the heavy use of statistical techniques in both analyzing and training bots(or fitting appropriate parameters). Which leads to a question. What kinds of materials afford the correct mathematical training needed for studying board game programming in particular go? It seems to me that you need a fair amount of statistical training but what kind? I have some background in statistics but it is quite limited so I am uncertain if this would make a good hobby- or if I need to "hit the books" so to speak. Thanks again, Carter. --- On Fri, 5/23/08, Jason House <[EMAIL PROTECTED]> wrote: > From: Jason House <[EMAIL PROTECTED]> > Subject: Re: [computer-go] batch size and impact on strength > To: "computer-go" <computer-go@computer-go.org> > Date: Friday, May 23, 2008, 1:47 PM > On May 23, 2008, at 4:22 PM, Magnus Persson > <[EMAIL PROTECTED]> > wrote: > > > Quoting Carter Cheng <[EMAIL PROTECTED]>: > > > >> I remember seeing this in the archive before but I > forget the > >> actual results of the experiment. Does processing > simulations in > >> groups of say 5-10 have any impact on the > strength of the program? > >> > >> Thanks in advance, > >> > >> Carter. > > > > Yes, at some point at least batch processing of > simulations will be > > inefficient. > > Every time you walk through the tree, the algorithm > prefers to go > > the best path possible to gain more information. With > each extra > > simulation in a batch there will be less and less > useful > > information. Maybe there is a tradeoff and you can get > away with > > small batches and gain something else. You have to > test and see what > > happens. > > I expect batching to help when multithreading > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
