Bot plays correct game through the end.

I expect that bots will improve.

Value of next moves is correlated to winrate calculated by bot.


When you correlate winrate with move value it
is obvious that MCTS bots are maximazing point difference
through all game untill endgame.
Than without normalization they only see 100*/* winrate
and cannot differentiate between moves which lead to win.

For move value with fixed R all count are win and winrate is 1.00,
so they do no look forward.

MCTS has some sort of semi-random normal distribution of outcome and
with normalization of this outcomes you get linear distribution which
is much better in the extremes.

I proposed how to determine Move value  so with normalization
bot can play best moves all game.

In example bot played all game as it was intended to play
through endgame. Like all MCTS bots try - to maximize point difference.
But current bots fail in endgame.
Now normalization window in example was fixed so it didn't fit exactly
and play is suboptimal. But not much.

Further when you get tree with point values you can use
all alfa-beta with prunning techniques (with statistics).

Now of course you can introduce bias, expert knowledge, null move
double move and similar techniques. Openings can be compiled ...


Leon.

_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to