[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] An [open] question on game tree search theory

Antti Huima wrote:
Erik van der Werf wrote:

>
>
> My idea about the whole thing is simply that some positions are
> difficult to understand (mostly because of tactical complications)
> without looking ahead, whereas other positions can be understood quite
> well without further expansion of the tree. Which positions require
> deeper search directly depends on the confidence you have in the
> evaluation for that position, and the probability that it will have
> any effect on your move selection in the root.
>
> Erik
>
>
Usually at this point somebody will suggest that evaluations should be
of the form "score + confidence estimate", i.e. that evaluations should
consist of two numbers, evaluation score, and some form of evaluation of
the confidence in the score. A generalization would be to let
evaluations be probability distributions of evaluation scores.

But in my opinion this just begs the question. In the game of go, maybe
the canonical measure for evaluation is the end result. Thus, if
evaluation is guessed (game-theoretic) end result, would "confidence"
mean some idea of how far the evaluation score is at maximum from the
game-theoretic end result? How do you quantify or approximate it? Can it
be quantified? What is "confidence in evaluation"?

In the game of go, pro players maybe win with only a few points margins.
Doesn't that kind of tight margin always end up being hidden in the
uncertainities of our evaluation functions during middle game?

In Magog I only distinguish between two types of evaluations.
1 Proven win/loss (like mate scores in chess). These normally do not need to be searched deeper.
2 Heuristic scores. These are searched deeper depending on some kind of likelihood that they will affect the root.

But I suppose more elaborate schemes can be used.

Erik

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/