It all started with perfect yose which current bots
play extremly strange.
My proposal of convering winrate to move value means
bot can play optimal move and wait for opponent mistake
or with lookahead of 2 or more moves it can try moves
which utilize small mistakes of the opponent. No insensible moves
like self ataris and waiting for big payoff. Same for winning
situations. Playing best moves shows good style. And with
small look-ahead you can prune moves with obvious refutations.
Good endgame means you can pose an endgame problem
and let bot play you perfectly. If you make mistake you will be panished.
If bot makes mistake than this is opportunity to learn the situation.
As I said before the object of the game is not only winning.
You learn from loses not wins. Collect wins against bots and you can
improve it.
Confidence and prunning are prety good as they are in MC but from
change of expected move points and status of groups you can deduce
strategical information.
As of window to determine winrate you use sliding
window. when winrates are close to 1 move window up un vice versa.
Same for size of window. And you pick random x for each playout.
Like for each counting you use random komi from window interval.
When you get winrates (for this window) for each next move you
basically get move values. So if bot is winning for 3,5 or 1.5 you can
see it
but with current play both moves get close to 1 winning rate.
I was very disturbed seeing bot winning for 0.5 point each game.
And best chess programs are exploring monte carlo methods as I recall.
Leon Matoh.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
