I see your point. First, I must confess I didn't read the entire paper you posted (too much like homework, and I'm still on vacation 
), but from what I read I don't see any reason this wouldn't work. But 3.3.2 seems to be the real issue with applying it to chess engines. That was just the KRK dataset. It seems to me that the amount of data that would have to be processed is prohibitively large. Please correct me if you think I'm off base; I always love discussing CS topics.
Theoretical Limitlessness of Computer Chess
Sort:
Hey all, I've been thinking about something these last few hours and I was wondering what your opinion was.
We all know that computers are totally dominating chess in almost every single way (though they do miss out on the best part -- the beauty of chess seems to elude them
), but the achilles heel of a program is its evaluation function-- the only thing those who denounce computers have working for them is that computers lack "understanding" of a position. This is completely true and will always be the case, but I've been thinking of an easy (perhaps the better word here is "implentable") solution.
A great paper (and frankly a very easy read, even for those not so computer-science incined (the parts explaining chess may seem a little patronizing :P)) by David Gleich called "Machine Learning in Computer Chess: Genetic Programming and KRK" discusses an important and innovative (though not at all new) idea called machine learning and genetic programming (GP). GP is a subset of machine learning where, much like in Dawinian evolution, a function "evolves" based on its past success -- overall, the program would gravitate towards a better implementation.
Do you think that if someone came along with a large enough database of games, we could reverse-engineer quasi-perfect play? I do not believe this is too outlandish, since after all just a few months ago the creator of Rybka concluded with almost complete certainty that the King's Gambit is, with best play, a draw. With this kind of processing possible, evolutionary chess doesn't seem so far off.
Here's a link to Gliech's paper: http://www.cs.purdue.edu/homes/dgleich/publications/Gleich%202003%20-%20Machine%20Learning%20in%20Computer%20Chess.pdf
And for those who didn't get the news, ChessBase's story on the refutation of the King's gambit:
http://chessbase.com/newsdetail.asp?newsid=8047