Author: James T. Walker
Date: 15:12:10 09/19/99
Go up one level in this thread
On September 19, 1999 at 15:12:34, Ed Schröder wrote: >>Posted by Enrique Irazoqui on September 19, 1999 at 14:30:25: > >>>The problem (book-learning alike) is that it hides the real strength of a >>>chess engine. These days it's not about the engine but more about >>>the extra elo you can gain by smart (and aggressive) book-learning >>>and opponent recognition will only make it worse. It operates hidden >>>hardly to see (notice) for the end-user unless you take a very deep >>>look in the system. >> >>It is like with learners. Once upon a time, it gave advantage to some >>programs. Not now. Programs either they have a learner or they escape >>with broad books. > >You must be kidding. Book-learning is dominant. Have a look at the Rebel >Century games on my pages. You will notice the match Rebel Century - >Fritz 5.32. Fritz 5.32 was caught 14-15 times on 1.b3 As a result Fritz >lost 14-15 times and therefore lost the match. Remove the doubles and >Fritz would have won this match. It's a perfect example how silly these >matches are for so called accurate testing. It is meaningless and misleading. ******************** The above paragraph bothers me a lot. I will try to find these games and have a look. The question that comes to mind is: Were you using Fritz at the point in time when it's book learning was broken? There was a point in time when it's book learning using auto232 was broken. This was fixed in the last SP1. Jim Walker ********************* > >>Opponent recognition is, as far as I can tell, something no program has. >>But if they all develop it, and this would be a genuine advance in intelligence, the >>engine will still have the last word. We are saying the same as 2 years ago >>about learners. > >See above. The current status of book-learning is far from being perfect. >If you want to top on all comp-comp events it still makes sense to put a lot >of energy in book-learning. I wonder if that is the goal of chess programming. > >You can not imagine what you can do with opponent recognition. In the end >the best "book-learner" + "opponent recognizer" software will top on all >comp-comp lists and can completely fail in manual games. What has this >to do with playing strength? You are testing software that has nothing to >do with the capabilities of the chess engine. > >>>>>This whole auto232 thing is so fragile that I can imagine people >>>>>don't want to touch it any longer. >>>> >>>>It has always been fragile, but more reliable than the very few manual games >>>>that can be played. Proof: you and I play thousands of automatic games, and >>>>seldom any manual ones. Why is that? :) >>> >>>It depends on the intention you are playing these thousands of games. For >>>me that is to improve Rebel as it gives me a lot of useful data. >> >> >>It is useful as the only way to get enough comp-comp games to make accurate >>quantifiable comparisons, in spite of the few problems here and there. > >Accurate? First step: play 300-400 60/60 games using all learned data of >previous matches. Second step: remove all learning. Third step: play 300- >400 games again. Good chance you see a 100 elo difference as I have seen. > >>Although I learn much more from one game I play manually than from 10 >>autoplayed games, but this is another question. > >Agreed :) > >Ed > >>Enrique
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.