Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: TDleaf from someone other than Baxter et al

Author: Stuart Cracraft

Date: 13:43:46 07/18/04

Go up one level in this thread


On July 18, 2004 at 02:40:03, Rémi Coulom wrote:

>On July 17, 2004 at 23:44:55, Stuart Cracraft wrote:
>
>>www.cs.ualberta.ca/~jonathan/Papers/Papers/td.ps
>
>If you are interested in papers on TDLeaf(lambda) from someone other than Baxter
>et al, you should read those by Beal and Smith. They invented the technique
>first, and applied it to chess. I find their papers to be more convincing than
>those by Baxter. They used self-play instead of online play, and played many
>more games. They managed to obtain weights that look much better.
>
>Two of their papers were published in the ICGA journal. They have papers in
>_Information Science_ (122, 2000, 3-21) and _Theoretical Computer Science_ (252
>(2001) 105-119), also. Unfortunately, I do not think that any of those papers is
>available online for free. The paper in _Information Science_ is particularly
>intersting because it introduces a technique called "temporal coherence" to
>speed-up learning by using an individual adapative learning rate for every
>weight.
>
>Another interesting reference that is available online is this technical report:
>http://www.cs.bris.ac.uk/Publications/pub_info.jsp?id=2000100
>
>Rémi

Remi -- I'm well aware of Beal's work pre-dating it and read it quite
a while ago.

However, it is Baxter et al who solidified and expanded. Don would likely
agree with this. It was in a far more limited domain of piece-value auto-tuning
that Beal and colleagues operated and at base positions rather than frontier
nodes, etc. as I recall.

And if you want to play the game of go-back, please go back to Art Samuel whose
work, I feel, anticipated Sutton's, Tesauro's, Baxter's, and everybody else's
in this area.

I was merely pointing out Shaffer's solid backing of TDleaf after having
spoke of it downwardly before -- he even alludes to that point in the paper,
so it's a big admission and step forward for a solid researcher and I
recommend that paper.

Manual tuning is for the birds,

Stuart



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.