Author: pavel
Date: 11:21:40 04/15/05
Go up one level in this thread
On April 15, 2005 at 13:23:46, Dieter Buerssner wrote: >On April 14, 2005 at 23:29:25, pavel wrote: > >>Also another thing I noticed, is that YACE has a limitation of 30,000 learned >>position. Is it true? I have already reached the limit. What happens now? > >There is a limit - yes. I actually forgot how large exactly (and cannot look it >up at this moment) - the 30000 might well be correct, although I thought it was >10000. It can be enlarged by some obscure commands - which I cannot look up >either at the moment. Perhaps you already have the limit enlarged? Didn't I give >you an explanation by mail how to do it some time ago? Do you remember José Luis >Jiménez? He should be able to explain it to you. IIRC I implemented it >especially for him and for you. I myself have done most of my tests with >learning disabled, so I never needed that many positions. I fear, that the >positional learning will not help that much in practice. But it helped against >the infamous Mercilious (sp?) attack on ICC, where Yace could lose very fast >many games in a row against a human. Also, Mogens Larsen had reported one case >agains Gandalf, where a lost variation was draw the next time, and a critical >move was avoided. Both opponents had in common, that they always tried hard, to >repeat a winning variation. At that time, Yace had no book learning. That should >take care of such a situation, too, and in practice might be more useful. But I >really have never tested carefully, how much the learning features help. > >Cheers, >Dieter I remember there was supposed to an update that would increase the learning limit. But I never recieved the updated engine. I still keep in touch with jose. As a matter of fact he recently sent me his updated learnpos.bin, and I checked to see that it was stuck at 30000. So it seems that he is also using the public Yace version that doesn't have the limit increased. I am using the PB version. I will ask Jose if he still has the workaround. I am preety sure I never recieved one. But we did discuss about it. I personally think learning helps alot in the long run. There are two ways we can find how much learning helps, one would be to check every positions learned and revisited manually for every games it learned, which is almost impossible. Another way would be to play large amount of games without any learning limits. Which is what I am trying to do. We can compare the result every few thousands games and see if the improvement curve goes upwards or not. IMO 30000 is a low limit. This is what my yace output says: --------------------------------------------------- Crafty is the magic word, Scid wants to see Yace (Yet another chess engine) version Paderborn Type help for a (very unhelpful) list of commands Make sure, you have a file yace.ini in the current directory, with a line: hash 30M or similar, otherwise Yace will only use a very small hash of 1M. feature done=0 DBBOOK101 400000 128 20 ff 455601 entries 87381 size 1048572 size_wanted 1048576 tts3 29126 pos_learn = 1 Stored 30000 learned positions into hash table book_learn = 4 DBBOOK101 10000 128 20 ff 10736 hash size wanted 80000000 Stored 30000 learned positions into hash table entries 6666666 size 79999992 size_wanted 80000000 tts3 2222221 5-men endgame table bases found book_learn = 6 DBBOOK101 10000 128 20 ff 10736 min_learn_tc = 60 white ( 1): ------------------------------------------------------------ Cheers, pavs
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.