[Computer-go] Computer-go Digest, Vol 17, Issue 78
dailey.don at gmail.com
Wed Jun 29 13:55:10 PDT 2011
On Wed, Jun 29, 2011 at 4:31 PM, Hendrik Baier <hendrik.baier at googlemail.com
> It sounds crazy to me that it works at all as it has no real knowledge of
>> the position.
> It sounds crazy indeed. But your typical MCTS (without node priors) has no
> knowledge of positions at all - it just learns which of the available
> actions seem to work best. The classifier actually has more knowledge than a
> typical MCTS tree, since it generalizes between positions, as you said. And
> it generalizes based on the intuition that good reactions to the same moves
> are often useful in many branches of the search tree, see our Power of
> Forgetting paper.
> I'm interested in integrating this neural network approach into MCTS such
> that the convergence properties of MCTS are not lost...
> By the way, as far as I understand the network is not built from scratch
> before each move, but before each game. Each move could be an interesting
> approach as well (maybe for longer time settings?).
Although I said "before each move", what I really understood was that it
was NOT pre-computed.
When I said it sounds crazy of course I wasn't being critical, I accept
that it is an interesting experiment.
Years ago I did some experiments with MCTS that tried to generalized and
played some games on KGS. The program was weak but I got some interesting
comments about it because several players commented that it seemed to
understand the positions well but sucked at technique.
I don't remember the exact details but I tried to treat patterns more like
moves while still using tree search. I have long felt that computer Go
needed better ways to generalize knowledge learned during the search and/or
> Computer-go mailing list
> Computer-go at dvandva.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go