Author: Uri Blass
Date: 20:59:42 01/14/01
Go up one level in this thread
On January 14, 2001 at 15:59:13, Howard Exner wrote: >On January 13, 2001 at 22:54:03, Dann Corbit wrote: > >>On January 13, 2001 at 21:44:45, Howard Exner wrote: >> >>>Here are some results by Tiger 13 and Century 3on an Athlon 900. >>> >>>Tiger 13 avoids Qe4 at 3:03 by choosing Ka6 at depth 13. >>>Later on the same ply it changes to b3 at 4:27 then b4 at 7:21. >>> >>>Century 3 has this output >>> >>>0:52 12th ply Qe4 +1.66 >>>2:18 13th ply Qe4 -1.29 >>> >>>Will not avoid the move Qe4 till the next >>>iteration in 14:22, playing b4. >>> >>>That seems like quite a long delay as the eval swung +2.95. >>>Given that Qe4 had -1.29 after 2:18 when it took an extra 12:04 >>>to resolve that b4 is better, why such a long delay? >> >>It's called fail low, followed by massive fail high. The program has to >>research the entire tree. A particularly bad combination. >>;-) >>I am sure that it would play the correct move as soon as the fail high occurs, >>if the time ran out. It just might have a faulty evaluation of the true value >>if it has to move early. That would get corrected on later plies. >> >>On crafty, for instance, when the fail high occurs, you will see >>++ <new move> >>but no score and it will just keep pounding until the time runs out or until it >>resolves the score correctly. When that happens, it has already found a move >>and knows that the move is better, but it does not know how much better or if >>there is an even better alternative [which is pretty rare after a fail high but >>definitely does happen]. > >Thanks. You can tell I'm not a programmer when I ask questions like these. >I was thinking that if a move was discovered to be bad after 2 minutes then >it should take at most 4 minutes to find a suitable replacement, for this >example at least as many times other moves are equally as poor. Another thought >that occured to me was that given the open position with Queens that Century >might have to resolve avoiding perpetual - this could explain the long think. I think that I can explain the long think. Rebel saw a big score for trading queens so it could get relatively big depth because trading queens simplifies the position. Rebel had only to prove that the score of other moves is lower than 1.66 and it is was relatively easy for it. After Rebel found a fail low it had to prove that b4 is better than 1.29 for black and it was hard for it to do it at the relevant depth. I think that the right way to solve the problem is simply to tell Rebel (only after a big fail low) to calculate an exact score of the second best move at smaller depth(for example depth 10) and if it finds a better move at this depth it can change the move that it wants to play to the move that it found and to investigate the same move at depth 13. Only if it finds problems at depth 13 it can change its mind again to the previous move. This idea should be used only after a big fail low. The idea may cause the program to play worse moves if it has not enough time to finish the investigation at depth 13 but there is no big difference between different losing moves. Uri
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.