Author: Vincent Diepeveen
Date: 09:44:12 07/02/03
Go up one level in this thread
On July 01, 2003 at 21:57:14, Dann Corbit wrote: >On July 01, 2003 at 18:28:19, Djordje Vidanovic wrote: > >>On July 01, 2003 at 15:35:22, Charles Roberson wrote: >> >>> >>> KnightCapp uses neural networks and I beleive some papers were published >>> about it. >>> >>> ICCA Journal vol 14 September 1991 has a paper "Neural Networks as a Guide >>> to Optimization" by A. van Tiggelen. It discusses automated tuning of a >>> position evaluator. >>> >>> Charles >> >> >>KnightCap is pretty good at learning both opening lines and evals. I finally >>managed to compile it to run well on my Linux box and I started it without any >>brain.dat actually and with default coeffs.dat. In a little while it started >>beating Gnuchess (say after about 400 blitz games) and even managed to win >>several games against Crafty. It IS getting stronger, but of course its limits >>are close because the search parameters have to be better. Although some >>programmers do think that its qsearch is fine... >> >>I find KnightCap a lot of fun, especially its "mulling" over the losing lines >>and finding out better lines while idle. I find it an extremely tough opponent, >>as tough, if not tougher than some well known programs. After about 500 blitz >>games it started beating Scidlet as well. Scidlet is considered to be about >>2200, so I'd venture to say that KnightCap progressed from a 1900 player to a >>2250 player in about 500 games. There are a couple of papers on KnightCap in >>pdf format, by Andrew Tridgell and Jonathan Baxter. And, of course, the famous >>paper on Temporal Difference learning by Rich Sutton that started it all. All >>can be found on the Net. >> >>Neural learning seems to be quite succesful in games that require fewer >>parameters than chess, such as backgammon (JellyFish, GnuBackgamon, BGBlitz) or >>checkers. Chess is tough as it has lots of parameters that matter. > >I think KnightCap is the most successful neural-net type of program. Lots of >programs have some sort of learning, but neural nets are not used successfully >very often. > >I think to work really well one would need: >1. A very fast computer >2. A very large number of evaluation terms >3. A very long time slot after a loss so that the calculations could be >absorbed into the eval. > >I know some other people working on neural nets, and I think the results are >interesting. > >The early attempts (e.g. Sal) were unfathomably lame. But then again, so are >the early attempts at anything and everything. > >The question is: >Can a neural net make a program play better? >I think the answer is pretty clearly "Yes!" I agree with you under the condition that you give it 10^120 computing time to optimize for chess. The problem is that todays search algorithms + evaluations are needing less to solve chess than a NN needs to just optimize for a few parameters. > >The more salient question is: >Is there a more effective way to make a program play better? >I don't think that question has been answered or even can be answered. > >That's just one of the things that makes chess programming so interesting.
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.