Petr Baudis wrote: > On Fri, Mar 07, 2008 at 12:43:42PM -0500, Don Dailey wrote: > >> I honestly think there are better ways to handle this, if you must, >> other than changing the goal to a losing goal. At least give the >> computer the right goal (winning) and adjust from there. >> >> If I were trying to solve this "problem", the solution I would look at >> first would be to pre-process the moves in advance of the search, and >> just impose a very slight bias on the root move list. The idea is >> that all other things being equal, it will play a patterned or "normal" >> move instead of a random move when the game is virtually over. >> >> If you do it right, you can probably make it play more like like it has >> an ego without making it weaker. >> > > Was making it play "nicer" moves anyone's goal here, actually? I always > wanted to make it stronger by dynamic komi, not just to make it play > moves that look nicer to humans. > The amount of improvement possible with this technique is pretty limited, so I think the primary goal is to make it play nicer moves. After all, it's already been admitted that the benefit, if it exists, comes from making it play better in losing positions, or in some cases treat the symptoms in such a way as the program may win some games for reasons it doesn't understand.
- Don > Of course, there is still the argument whether dynamic komi would help > this goal against human players, and I have nothing new to bring into > that discussion. But the bots playing ugly moves is not a big issue for > anyone, I believe; also the stronger the UCT bots get, the less silly > looking moves they seem to play even when the game is over. > > That's probably because the stronger playing bots impose some move ordering on the program in the hopes that strong lines are considered early based on a priori knowledge. This has the side effect that those same moves, even if losing, have a better chance of being the move selected by the program in cases where there is little difference between moves. Even my simple and weak program improved somewhat cosmetically when I did something similar. - Don _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
