The utility of the score isn't linear.
But you have a good point that the score should be more important when
the winrate is more extreme.
There have been somewhat successful attempts to score with a "lazy s"
function.
It might prove fruitful to make the parameters of this function
dependend on the winrate.
(Don't ask me which winrate. Perhaps the winrate of the last move, or
dynamically changing on the current winrate)
Stefan
This thought was inspired by the dynamic komi discussion
(specifically, Lewkaz's comments) and it may be most appropriate in
handicap games, if appropriate at all. I was wondering if anyone has
tried it. The idea is to replace the WinRate term of your move
selection (at the root and/or in the tree) with a linear combination
of WinRate and ExpectedScore in such a way that WinRate matters more
when WinRate is near 0.5 and ExpectedScore matters more when WinRate
is near an extreme. Something like this:
extremeness = 2 * abs(WinRate - 0.5)
Eval = (1 - extremeness) * WinRate + extremeness * ExpectedScore;
I think ExpectedScore would need to be in the range 0 to 1, just like WinRate.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
