Author: James Swafford
Date: 14:27:52 06/02/04
Go up one level in this thread
On June 02, 2004 at 16:58:01, Dann Corbit wrote: >On June 02, 2004 at 16:07:14, James Swafford wrote: > >>On June 02, 2004 at 16:03:10, James Swafford wrote: >> >>>On June 02, 2004 at 10:06:22, Dann Corbit wrote: >>> >>>>On May 29, 2004 at 16:13:24, James Swafford wrote: >>>> >>>>>On May 29, 2004 at 14:15:37, Frank Phillips wrote: >>>>> >>>>>>On May 29, 2004 at 12:53:54, James Swafford wrote: >>>>>> >>>>>>>On May 29, 2004 at 11:53:29, Frank Phillips wrote: >>>>>>> >>>>>>>>On May 29, 2004 at 11:44:58, James Swafford wrote: >>>>>>>> >>>>>>>>>On May 29, 2004 at 11:35:07, Frank Phillips wrote: >>>>>>>>> >>>>>>>>>>On May 29, 2004 at 04:00:31, Gian-Carlo Pascutto wrote: >>>>>>>>>> >>>>>>>>>> >>>>>>>>>>>I don't think so. The program still has weaknesses that a bit of >>>>>>>>>>>extra hardware will not overcome. >>>>>>>>>>> >>>>>>>>>>>GCP >>>>>>>>>> >>>>>>>>>>What are these weaknesses? >>>>>>>>>> >>>>>>>>>>Bob may even be able to fix them before the event. >>>>>>>>> >>>>>>>>>He was talking about his program, not Crafty. >>>>>>>> >>>>>>>>Thanks. I misread the post. >>>>>>>> >>>>>>>>But I am still interested in the weaknesses being referred to by GCP, which are >>>>>>>>resistant to faster hardware. I have so many myself. If only I knew what they >>>>>>>>were :-) >>>>>>> >>>>>>> >>>>>>>As in, "I can't seem to mate Shredder, even with faster hardware!" ?? :) >>>>>>> >>>>>>>-- >>>>>>>James >>>>>> >>>>>> >>>>>>I guess the answer is yes, although I have never had better hardware - and am >>>>>>not SMP, so probably never will. >>>>>> >>>>>>See you tonight at ICC author's only tournament ? :-) >>>>> >>>>>NOt as a competitor-- my thing is nowhere near strong enough >>>>>to compete yet. I'm hoping to be able to compete in the next >>>>>CCT, though. >>>> >>>>Are you still doing the learning stuff? >>> >>>I've been working with TDLeaf quite a bit. At some point I'll >>>post something with some meat to it, but to sum it up, I'm >>>not nearly as optimistic about it as I once was. >>> >>>In my experience, TDLeaf can train the material weights, and it >>>can even produce an evaluation vector that's superior to a >>>'material only' vector. I am not convinced it's useful for >>>training a complex vector, nor am I convinced it does a better >>>job than hand tuning. For that matter, I am not even >>>convinced it converges to the optimal vector! >>> >>>Caveat: it's possible (though I think it's unlikely) that >>>my implementation is flawed. My engine will become open source >>>at some point (maybe after the next CCT), so you can judge >>>for yourself then. >>> >>>Will Singleton and I had a bet on this... I conceited defeat >> >> >>Gah! I "conceded" defeat. >> >>>the other day. THe original bet was for the loser to fly >>>the winner and spouse across country for drinks. :) I'm >>>pretty sure Will's decided he'll forego that if I show up >>>at a tourney, but that's his call. >>> >>>I'm still very interested in learning algorithms, but I'll >>>be focusing on improving my evaluation for a while. >>> >>>Again- I will post some data at some point. > >I am doing a computer guided optimization for Beowulf. > >It takes ~12,000 positions from super-GM games and SSDF games among the top >computers where all the participants chose the same move (no other moves chosen >for that position). > >For each of about 100 parameters, I vary the value from too small up to too >large (e.g. a knight might go from 200 centipawns to 450). At some optimal >point, the largest number of positions will be chosen. I fit a parabola >throught the data ans solve for the maxima (if any). > >Often, the variance of the parameter has no effect on the solution scores (for >instance, I might get 5500 solutions no matter what the parameter is, or the >number of solutions may vary randomly). So I also solve for the minima of the >time curve. As an example, a depth 4 search using NULL MOVE will probably solve >a few LESS positions than not using NULL MOVE, but it will take 1/3 of the time >at some optimal prune level. > >I have had lots of bugs in my curve analysis, but I am slowly working it out. > >Before, I solved for a smaller subset of tactical positions which made it great >at solving those tactical positions but lousy at playing. I am hoping for a >better result this time (especially since some of my result calculations were >backwards, making the fits enormously unstable). That is fascinating. Would you explain in more detail how you went about choosing the positions? I see extracted positions from very high quality games, but I don't understand what you meant by "all participants chose the same move." Do you mean that you (1) took a position (and the move made in that position) from a game and then (2) looked in other games for the identical position, and if found (3) compared the moves, keeping the position if the moves matched? IN the case of the knight that might vary from 250->400; how much would you increment for each run? 5? 10? It seems you need to watch for local optima. Perhaps you find "optimal" for parameter A, then move on to find "optimal" for parameter B, but perhaps some other combination of both would've done even better. If you extend that out to all combinations, the number of possibilities is truly staggering. Any thoughts on that? Maybe a random mutation here and there? Your time curve analysis is spot on: very clever. I learned in early TDLeaf experiments with Prophet that if you get crushed tactically (as Prophet usually did at the time), you begin to train the eval to predict tactical blunders. So- I improved my search a bit and tried again. I think part of the problem now is that my eval is too simple. I deliberately kept the eval very simple, thinking I could do a "proof of concept" type thing, then just start throwing terms in, letting TDLeaf train the weights. Well, what you start to see is outrageous values for some parameters to compensate for missing knowledge. Perhaps such 'outrageous' values could actually serve to point out exactly what's missing, but I don't know how yet. Anyway, I think parameter training of some flavor will be one of the next 'big things'. What would be _really_ cool, though, is learning not only how important a term is, but finding new terms! (i.e. pattern recognition and all that) Good luck with your project: it sounds like a sound idea... :) -- James
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.