Author: blass uri
Date: 16:05:53 10/19/98
Go up one level in this thread
On October 19, 1998 at 18:12:40, Robert Hyatt wrote: >On October 19, 1998 at 17:22:10, blass uri wrote: > >> >>On October 19, 1998 at 16:18:41, Robert Hyatt wrote: >> >>>On October 19, 1998 at 12:03:15, blass uri wrote: >>> >>>> >>>>On October 19, 1998 at 10:14:42, Robert Hyatt wrote: >>>> >>>>>On October 18, 1998 at 14:25:23, blass uri wrote: >>>>> >>>>>> >>>>>>On October 18, 1998 at 13:41:29, Robert Hyatt wrote: >>>>>> >>>>>>>On October 18, 1998 at 12:13:34, Alessio Iacovoni wrote: >>>>>>> >>>>>>>>1) Shouldn't computer strenght it rather be measured on "average" entry-level >>>>>>>>computers.. i.e. the ones actually used by the majority of people? >>>>>>>> >>>>>>>>2) Also.. do programs benefit in the same way from higher speed and increased >>>>>>>>hash tables? If not, tests would not be comparable, therefore useless. >>>>>>>> >>>>>>>>3) Why are books used in tests? Shouldn't a top level computer program be >>>>>>>>capable of doing at least decently in the opening phase *without* resorting to >>>>>>>>it's book? If the answer is no.. then it could be easily beaten by even >>>>>>>>lower-performing computers by having it systematically go out of book. Or am I >>>>>>>>wrong? >>>>>>> >>>>>>>Computers would do just as well without a book as a human that had *never*seen >>>>>>>an opening book. And I'd bet the human would fall into many of the same sorts >>>>>>>of "traps" that the computer would. But even worse, the computer would tend >>>>>>>to play the same opening every time, since the tree search is deterministic. >>>>>> >>>>>>There are some variable in the evaluation function that you can decide that they >>>>>>will not be constants >>>>>> >>>>>>For example suppose you have a positional bonus for a pawn in the 5th rank of >>>>>>0.2 pawn. >>>>>>You can decide that the positional bonus will be different(0.23 pawn or 0.17 >>>>>>pawn) >>>>>>You can decide before every move to change the positional bonuses by a small >>>>>>random number and it may cause the program not to play the same opening every >>>>>>time. >>>>>> >>>>>>Uri >>>>> >>>>> >>>>> >>>>>Sure... but it can also make it play *weaker* in addition to playing more >>>>>random... >>>> >>>>I do not think that more than 20 elo weaker if you change only by a small >>>>number(every positional bonus will not be changed by more than 0.03 pawn). >>>> >>>>I do not think that the positional bonuses are optimal(I think that noone knows >>>>and by doing games you can get get closer to optimal) >>>>Another problem is that the optimal bonuses for Blitz may be different from the >>>>optimal bonuses for slower time control. >>>> >>>>I think that for slower time control it may be better to increase the positional >>>>bonuses but I am not sure about it. >>>> >>>>Uri >>> >>>The problem is that "20 Elo" is misleading. When you are talking about >>>computer vs computer matches. A small change in one program often produces >>>a big change in the match results, because that becomes the *only* thing that >>>separates the programs... IE If I test two crafty versions that are identical, >>>but let one use 1 cpu and the other use 2 cpus, I get huge margins of victory >>>with 2 cpus, yet when I play that same test match against humans, the two cpu >>>version will score somewhat better but not nearly so much as the crafty vs >>>crafty match suggested... >> >>I think that there is also a difference between crafty vs crafty matches and >>crafty vs other programs matches. >> >>Uri > > > >I'm not sure I understand your comment, because obviously crafty and other >programs are different. But I try to tune the eval terms to play the best >chess I can make it play. Adding random "fuzz" will only weaken it, not >help it... A programmer may play many games against crafty without telling you and do an opening book from the lines that the program won. If you add random "fuzz" then you can escape from the opening preparation. 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.