Message from discussion
Coevolution as game tree search
MIME-Version: 1.0
Received: by 10.100.119.17 with SMTP id r17mr195455anc.11.1214465395861; Thu,
26 Jun 2008 00:29:55 -0700 (PDT)
Date: Thu, 26 Jun 2008 00:29:55 -0700 (PDT)
In-Reply-To: <D80B6C29-C9A6-47E8-8EAF-983B1223004E@lclark.edu>
X-IP: 139.230.82.60
References: <D80B6C29-C9A6-47E8-8EAF-983B1223004E@lclark.edu>
User-Agent: G2/1.0
X-HTTP-Via: 1.1 SCISMLFW1
X-HTTP-UserAgent: Mozilla/4.0 (compatible; MSIE 7.0; Windows NT 5.1;
InfoPath.1; .NET CLR 2.0.50727; .NET CLR 3.0.04506.648; .NET CLR
3.5.21022),gzip(gfe),gzip(gfe)
Message-ID: <4ab65af5-4bf0-4f9e-bd5c-9eb5c41b1656@g16g2000pri.googlegroups.com>
Subject: Re: Coevolution as game tree search
From: phi <PhilipHings...@gmail.com>
To: Computational Intelligence and Co-evolution <coevolve@googlegroups.com>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable
Hi Peter
I had a look at your paper at http://webdisk.lclark.edu/drake/publications/=
drake-gem2008-final.pdf
=2E Interesting. Haven't seen this idea before (but that doesn't mean to
say it doesn't exist). It seems vaguely reminiscent of LCS's, but not
the same, and as you point out, a lot of these things (GA's,
reinforcement learning, LCS...) can be described as sort of random
search with a bias towards things that are learned to be good.
Certainly seems like the idea could be applied to other similar tasks.
cheers, phi
On Jun 23, 12:18=A0pm, Peter Drake <dr...@lclark.edu> wrote:
> Greetings, everyone.
>
> My research group is exploring coevolution as game tree search, =A0
> specifically for the game of Go. We are aware of many attempt to use =A0
> genetic algorithms to evolve players or board evaluators over the =A0
> course of many games, but we are trying to use coevolution to find =A0
> the best move (or at least a good move) during the course of a single =A0
> game. The individuals being evolved are partial strategies, i.e., =A0
> subsets of the game tree starting at the root but not extending to =A0
> the leaves. When we have two individuals play against each other as =A0
> part of fitness testing, we use the individuals to determine moves =A0
> until we fall of the bottoms of the trees, then finish the game =A0
> randomly, as a Monte-Carlo playout. We've had some preliminary =A0
> success, and will be presenting a paper at GEM'08 next month.
>
> My question: is anyone aware of other work in this area? We haven't =A0
> been able to find any.
>
> Thanks in advance,
>
|
