Author: Robert Hyatt
Date: 15:58:46 07/27/98
Go up one level in this thread
On July 27, 1998 at 13:36:57, Don Dailey wrote: >On July 27, 1998 at 05:35:21, Robert Hyatt wrote: > >>On July 26, 1998 at 19:57:24, Don Dailey wrote: >> >>>----------- SNIP --------- >>> >>>>>That's how I feel about it. In chess it takes two to tango, >>>>>the computer gets a chance to control the play too. >>>>> >>>> >>>> >>>>while this is true, the computer has only one "tango" it can dance to. It >>>>won't self-modify as the games are repeated. It will learn to avoid losing >>>>openings, but it won't learn to avoid losing "plans." I could show you about >>>>500 different ways to transpose into a stonewall. Roman ragged hell out of me >>>>about this until I simply decided to "teach" crafty about the stonewall system >>>>directly. I've watched him do this to many programs also. And all he has to >>>>do is to find different transpositions, even if he gives up a tempo or two >>>>to accomplish it. >>> >>>I am glad you are finally comming around to my point of view on this >>>one. When you say, "even if he gives up a tempo or two to accomplish >>>it" you are hitting at the heart of the matter. And I don't care >>>in the slightest if Roman can still beat the hell out of it. The >>>point I DO CARE about is that Roman feels his best plan against the >>>computer is to give up a couple tempo's to get a stonewall attack! >>> >>>I don't know about you, but if I didn't understand an opening very >>>well I would be happy to be able to play it a couple tempo's up. >>>If I was playing a grandmaster who was very likely to beat me in either >>>case, I would still rather have the extra tempo's, but that's just >>>me. And I would take some consolation in the fact that my opponent >>>avoided objectively better lines to get me into this. It wouldn't >>>matter to me whether he did this out of contempt for my ability >>>to play that opening or out of respect for my ability to play >>>the other lines, in either case I still have the advantage of >>>the extra tempo's. >>> >>>- Don >> >> >>However, the problem is, the tempi mean *nothing* for black in that opening, >>*if* black lets white achieve the "stonewall position". Black has very few >>options *after* the fact. And if he "fights" *before* the position is set up, >>white doesn't have to give up much at all. >> >>My point with this is that the GM's (and IM's) figure this out after a few >>games. And book learning doesn't help. And since programs don't learn anything >>else (except for some primitive "position learning" that doesn't help) they fall >>into this over and over. Or else you do as I do and try to recognize the >>pattern and penalize it. >> >>But until a program reaches the stage where they can "get burned" a few times, >>isolate the common "features" and then start avoiding them, a GM has a big >>advantage. The most common plan on ICC is to try an odd first move or two, >>just to get a program out of book. Because once you can do that, then you can >>go "fishing" to find plans to beat it, knowing full well that it won't be >>paying attention and parry where you "feint"... Any good master will learn how >>to take advantage of this... hence the famous "trojan horse" opening that >>killed so very many programs. Ed posted something about getting caught by it >>himself here a while back. > >I used to beat Fidelity programs this way. I could almost always >win against one of them because I took to the time to find all the >pathways to a win, which because of the variety book required >me to do this many times. Sometimes even after the opening was >over I would face a point where the program varied depending on >timing factors I assume. I didn't find the complete pathway, >I usually would just find a way to get a good position where I >could beat it most of the time. time out for a funny story. Back at USM, Bert Gower and I were regular members of the USM Chess club, and we had 4 players that I would consider "strong" chess players, mixed in with some sort of average players. The 4 "contenders" for every tournament were myself, Bert, Larry Mead, and Mark Zaremba. We started a tournament once where we would play one game per week. We had an odd number, so Bert suggested, and everyone agreed, that we enter his "SuperCon" to even things out. The funny part is this. Larry was a 2200 player, and Bert and I were not quite up to beating him very often. So he sort of "expected to win." In this case, Bert and I had an ace in the hole, because we found, quite by accident, a really bad "horizon effect" position where we could win a piece. I played SuperCon first and beat it. Bert played it (and, due to luck,had the right color and played basically the same game, at least to the point where he won the piece) and he won too. Larry had to play it, didn't know about our "trick" and lost. I won the tournament, Bert finished second. Larry finished pissed. :) Then we told him. He was *really* pissed. :) > >This is a strange situation because I really was not as good as >it was, and yet I rarely lost after the initial learning experience. > >This shows that a very important component is missing from our >programs, the ability to learn from this and to adjust. We do >it in little ways with the book and with the hash table learning >thing but it's not quite the same. These two things together >would probably have made it harder for me to beat fidelity the >way I did but cannot help much on a higher level, such as finding >more general methods to beat it. > >- Don I agree. Our programs do learn, but they do it second-hand, by our looking at the games and modifying the code to fix the holes. Left alone, they are all terribly vulnerable... And ICC shows *just* how vulnerable when you leave one night and your program is 2800, and when you return the next morning and find it at 2200. :) Used to happen a lot, but hardly ever any more. But holes can be found, and then exploited...
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.