> > I'm sure many of us are surprised how well this stuff > works. > > I'm not surprised because I knew a little about the principle 10 years ago. I create a game based on English/British checkers but played on a 6x6 board and a slightly different jump rule (you can only jump one piece in a move.) It was just for testing idea on searching and I wanted to keep it simple.
I discovered that a monte/carlo based evaluation function is superior compared to a simple hand crafted one. It was better even at the same time control. This game was small enough that I could search incredibly deep and discover diminishing returns (which must happen in all games but happens much more gradually than intuition would suggest.) I tried a naive version of this with GO and discovered that it could easily beat my alpha/beta searcher which had a naive static evaluation function (because I had just learned the rules and didn't understand the strategy whatsoever.) But it was still very weak and I temporarily lost interest in it. Then I saw a paper on the web about "gobble", a go program that evolved a move list ordering using simulated annealing. My interest was rekindled, I continued to tinker and eventually produced some simple MC based programs which were predecessors of AnchorMan. During this period other papers started to appear which I followed closely and now it's a common technique. Searching randomly is probably the quickest way to get an overview of the landscape. It's not very methodical, but methodical methods are much slower at quickly discovering the big picture. I'm thinking that the next big leap in computer go will be based on generalizing the knowledge gained from the play-outs. Right now, we structure this knowledge into a tree which is certainly an appropriate representation but we are still throwing away a lot of useful (but fuzzy) knowledge about the positions. (For instance we have to discover over and over than a certain move is good, not just in one line but probably in most lines of play.) There has been a lot of progress in this direction but I'll bet there will be more. I have tried some ideas that were failures so far - but I think this is a productive direction in general. - Don > _______________________________________________ > 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/
