Author: Michael Yee
Date: 18:26:56 04/28/05
Go up one level in this thread
On April 28, 2005 at 16:20:24, Jay Scott wrote: >On April 27, 2005 at 17:57:10, Michael Yee wrote: > >>Okay, so more work could be done here. I'm considering using very large >>(hopefully representative) training sets, not just self-play (as long as I can >>find the computing time!). > >In self-play, mistakes may go uncorrected. If the program does not know how to >punish a given mistake, it will make the mistake and never notice. (Backgammon >programs that learn this way escape the danger because the dice rolls force them >to explore different situations.) > >But collections of grandmaster games are also poor. Grandmaster moves are >produced by reasoning that a program cannot hope to approximate (just as program >moves are produced by reasoning that a grandmaster cannot hope to approximate), >so the lessons that can be drawn from them are necessarily shallow. There is no >such thing as a "representative" training set of human-produced moves or >positions, and machine learning algorithms are easily misled. > >I think the problem of finding good training data is unsolved. In fact, I think >that that is the central problem: We have good learning algorithms that can fit >data accurately and avoid overfitting. We do not have good data. > > Jay It's true that humans and programs use different techniques for choosing moves. But I'm still interested in trying to approximate human strategies with either (a) existing chess programming techniques (extensions/reductions/qsearch/static eval) or (b) some new yet-to-be-discovered approach. Path (a) seems questionable since the class of models almost surely doesn't contain the human strategies. But given this unavoidable "structural" error, it'd still be interesting to see how well we can estimate parameters for it. (I seem to remember deep thought or deep blue using some comparison training to tweak parameters somewhat successfully.) On the other hand, path (b) will require some kind of breakthrough. Maybe even something along the lines of Jeff Hawkin's theory of how the brain works: http://www.onintelligence.org/resources.php My dilemma is that if I want my research to fit under the theme "human heuristics for decision-making", I'm kind of constrained to try to "learn" in some way from human games. So even if I agree with your point about questionable training data (which I do to a large extent), I might be stuck with it. Michael
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.