I've been trying to think about what I really want and how it would be achieved.
What I'm interested in is the ability for bots to assign an accurate final score to a game as early as possible. When trying to think how to make this into a competitive challenge, I tried to consider both bots submitting votes and then finishing the game to prove who was right. Let's say bot A says it won by at least 10 points and bot B says it lost by no more than 7 points. The bots could then finish the game and see who was more right by having each bot continue the game to a completely clear end position. This scheme seems flawed when the ranges for the outcomes overlap. I was able to create sandbagging strategies that makes that style of competition less interesting. Here's an alternate competition that could work reasonably well. For now, assume a reference game is given. Bot A can declare an outcome of the game (say black +6.5). Bot B can then challenge the outcome by declaring the side that can do better. Bot B then plays the side that can do better and Bot A plays the side that supposedly can not. Both bots then finish the game and a winner is declared. Let me give an example: Reference game is given Bot A concludes black + 6.5 Bot B concludes black + 7.5 Bot A declares the outcome as black + 6.5 Bot B challenges the outcome and must win by more than 6.5 Bot A takes white and can not lose by more than 6.5 The game is played until the end and the outcome is 6.5 Bot A is declared the winner. If such a competition existed, would others be interested in competing? On Nov 6, 2007 10:48 AM, Jason House <[EMAIL PROTECTED]> wrote: > On Nov 6, 2007 10:30 AM, Lars Schäfers <[EMAIL PROTECTED]> wrote: > > > By the way: a 9x9 CGOS server using japanese rules... I have a dream.. > > ;) > > > > It's not a bad dream. That kind of thing could help spur development of > good ways to handle Japanese scoring. I fear that programs which are weak > would have a really tough time with Japanese scoring. > > I personally put a lot of weight on making my bot playable with people. I > have ignored this aspect of my bot since going MC since I've been doing a > lot of rewriting and getting basic functionality. Now that my bot has > achieved ~1400 ELO, however, I'm likely to start working on this human > element again. > > If we put up a milestone like 9x9 Japanese CGOS starting 1 Jan 2008, I'd > be very likely to upgrade my bot to handle it in prep for that. The delay > is mostly so that I both have time to implement it and to work in a few > other major features such as those used by MoGo, CrazyStone, and greenpeep. > If I'm lucky, I'll be over 2000 ELO this time next month... Out of > curiosity, how well do those strong bots do when limited to 10k playouts? >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/