[extropy-chat] computer chess again

[Alejandro Dubrovsky]

On Mon, 2003-11-10 at 04:22, Chris Hibbert wrote:
> This makes some sense to me.  I don't follow chess, but I do pay 
> attention to serious backgammon.  There are several extremely good 
> backgammon programs these days (all available for reasonable prices if 
> you don't mind running them on a PC.)  They all seem to be based on 
> neural nets, and have a fairly consistent style.  They are able to 
> report on their decisions on a play-by-play scale, and can give advice 
> about particular moves.
> 
> The best players are starting to come to understand, after very 
> detailed analysis, the choices the programs are making in some 
> non-obvious cases.  They are starting to incorporate the unconventional 
> strategies and trade-offs in their own games, and I've seen at least 
> one book that tries to explain to mid-level players what the experts 
> are learning about backgammon from the expert programs.
> 
> It wouldn't surprise me to hear that chess players, with their long 
> history of studying the games of other experts, are taking lessons from 
> the grand master programs.  The best human players should be able to 
> keep up with the best computer players, except in exhaustive cases like 
>   the three piece end-game puzzle that takes 147 moves to mate.  It's 
> also not surprising to find out that perfectly played chess should 
> usually lead to a draw.
> 
This is not happening much in chess mainly because humans cannot learn
to see 8-9 moves ahead at close to full breadth search like computers
do, and grand master humans are still kilometres ahead of computers when
it comes to evaluation of the board, and strategic planning (this does
not mean that games are not analysed with computers.  It's likely that
no GM analyses a game without one, but this is done for detecting
tactical holes, not for strategic planning on what the idea should be). 
Note that no successful computer program does extensive use of neural
networks since this would make the evaluation too slow, and the program
wouldn't be able to search many nodes.  This is not much of a problem
for backgammon because it is played in a more strategic manner, with
calculation either only being a couple of ply deep (or in the endgame),
since the possible moves in backgammon at each ply are MUCH bigger than
in chess, making proper calculation impossible for humans (and not very
deep for computers).  Backgammon is a very special case for computer
game playing since due to this very shallow depths attained it's very
hard to write a good evaluation code, and because of its probabilistic
nature, a standard mini-max searcher would be too cautious. It was the
first game where self-learning neural networks learnt to play the game
at a master level (the original TD-gammon had no calculation at all
AFAIK, so it's pretty good achievement for NNs.  Later versions had up
to 3-ply depth (a ply is a half-move btw) and maybe other programs
calculate deeper, but depth doesn't seem to be very important in
backgammon since it doesn't seem to improve the program much.  A chess
program that saw three ply deep vs the same program when seeing 2 ply
deep would just kill it)

> Are chess players publishing analyses of the games of the top-level 
> programs?
> 
Public GM vs computer games are published and analysed.  But there are
very few of these as there is very little for the GMs to win.  99.9999%
of the real GM vs computer games are private of course.

alejandro