Author: Steve Coladonato
Date: 09:52:29 02/24/00
Go up one level in this thread
On February 24, 2000 at 11:35:22, Laurence Chen wrote: >On February 24, 2000 at 11:02:28, Steve Coladonato wrote: > >>Given a position, let's say at move 19 for white, the computer program will >>evaluate it and return either a numeric or symbolic evaluation and a variation >>that is considered best for both sides. Also, let's say that at the position >>for the next to last move in the variation that is returned, say move 22 for >>white, that in actuality four ply beyond that, move 26 for white, there is a >>refutation of the complete variation. The outcome being that the suggested >>variation is not good at all. This would probably be discovered at move 21 for >>white, but in the interim, the player is being directed down a bad line. Given >>all that, my question is does this really happen in an evaluation of a position >>and if so, how is it handled? I can understand that if the program "learns" >>that it probably would not choose this variation again. But even then how far >>back does it need to go to correct the line? >> >>Please disregard errors in the ply count. >> >>Steve >I've yet to see a program learn from this. All programs suffer from this >problem, it is known as the horizon effect. Do you know any program which >learns from this? I guess the easiest solution is to edit the opening book, and >mark the variation as bad so that the engine will completely avoid the position >before it arrives to that crucial position. On the other hand, if you set up a >position then how will the engine learn from it. From a human perspective, it >can be a trap for those who depends too much on computer evaluation, and then >the human player complains that the engine sucks because it mis-evaluated the >position completely. Hence, there's still a lot of improvements to made for >chess engines, and they still got a long road to climb. >Laurence I can understand what to do if the condition happens while you are still in the opening book, but I'm mainly interested for the condition occuring in the middle game. I haven't seen the commercial programs "learn" from this, but I'm not strong enough to beat them at sensible time controls. I have seen the Sapphire II learn from a losing game, but it doesn't change it's move order early enough in subsequent games. It still works itself into a hole only a little differently. Steve
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.