Computer Chess Club Archives




Subject: Re: chess and neural networks

Author: Sune Fischer

Date: 05:48:10 07/02/03

Go up one level in this thread

On July 02, 2003 at 08:43:43, Djordje Vidanovic wrote:

>On July 02, 2003 at 05:07:19, Sune Fischer wrote:
>>I'm not sure nets are really that good for chess.
>>First of all you would need a huge NN with the capacity to remember of all these
>>patterns, I mean thousands probably tens of thousands of neurons and weights.
>>This would make it very slow to compute at runtime, if not using special
>>designed hardware.
>>Secondly, even if one could design a good cost function, like I suppose in
>>KnightCap, what would the network learn?
>>Well, it might learn double pawns tend to be bad, it might learn 2 rooks is
>>better than 1. We already know this of course, and I think it's going to be hard
>>to train the network to do better than handtuning here.
>>The only interesting aspect would be, if the NN could teach itself some rules we
>>where not aware existed. Honestly I don't think that is very likely, what is far
>>more likely is that the net learns something that is wrong.
>>When it's working for humans, I think it's because we litteraly do have billions
>>of neurons, and it's not that slow because it's running in parallel.
>>In spite of this our evaluation is nowhere near perfect, we aren't all world
>>champions and we still need to search for tactics.
>I think that KnightCap has a rather good neural net setup.  The latest score,
>played at 2 min for 40 moves, against The Crazy Bishop finished 35-19 for TCB.
>We all know that TCB should be about 2400, even higher at fast time controls, so
>I think that KnightCap must be rated now about 2300.  Pretty good for a proggie
>that was kicked around some 500 odd games ago.   I am not saying that NN will
>solve the problem of evals and be the crucial breakthrough in computer chess
>programming, I am only stating that NN does have a point and that, perhaps ( a
>big question mark here), just perhaps, some sort of NN learning could be
>implemented in future chess engines.  It is incredible to watch some of the
>games it which KnightCap literally wiped out TCB using merely one third of the
>allotted time...  Lots of fun, try it, you will appreciate the program, I am

I agree it is very exciting to study this thing :)

I wasn't aware they used a NN in KnightCap, I thought they used a standard eval
with some kind of board control array and then TDLeaf to train it.

Maybe I should have another look at it :)


This page took 0.01 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

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