Author: Miguel A. Ballicora
Date: 20:53:46 12/27/01
Go up one level in this thread
On December 27, 2001 at 19:08:45, David Rasmussen wrote: >On December 27, 2001 at 14:13:01, Robert Basham wrote: > >>Greetings! >>Have you seen this link of basis chess test problems with some computor results? >> >>It looks very good to me. >> >>http://www.multimania.com/albillo/cmain.htm#tests >> >>Robert > >I know if these tests. They are good, if a little old. But they are really not >what I was thinking of. For at least two reasons: > >1. They are "difficult" tests that tests advanced knowledge >2. Many of them isn't a test of knowledge so much as search depth. > >I want tests that are like this: If the program doesn't make the prefer the >right move, instantly, you have a problem because it indicated that some basic >evaluation term is over- or underevaluated. Like the program doesn't prefer some >move, because it doesnt understand that two connected passers are strong, or it >doesn't prefer some move because it doesnt understand that an outpost is strong, >or that owning an open file is strong etc. Really basic evaluation terms. This >would test two things: > >1. That the program has a certain (more or less basic) evaluation term necesary >to recognize the advantage of the position obtained by playing the key move >2. That the evaluation terms are more or less correctly calibrated in relation >to eachother, since it doesn't prefer some other move because it overevaluates >some other term or underevaluates the term in question. > >That means that the best positions would be ones that had several imbalances, >that could mislead the program. > >What I want is positions where you can say "if your program prefers to double >the rooks in this position or move the king to safety or something entirely >different instead of the 'right' move obtaining a strong knight outpost, your >evaluation terms are miscalibrated". > >I know that evaluation is subjective in some sense and that such positions and >their key moves would always be subjective in this sense. Even so, a testsuite >of such positions could really get one started on fixing the worst and most >basic evaluation errors of a program. I think that you mean a position like this [D]8/p2kbpp1/1pn4p/4p3/4P3/2N1BP2/PP3KPP/8 b - - bm Bg5 where a program should play Bg5 almost immediately if it knows about good and bad bishops and control of knight outposts (d4) in this case. I plan to build a test like this, but I haven't had the time yet. This position is from Jeremy Silman "Reasess your chess". Regards, Miguel > >/David
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.