Author: Bruce Moreland
Date: 10:01:47 04/10/99
Go up one level in this thread
On April 10, 1999 at 11:39:45, Keith Kitson wrote: >I don't think it is going to be quite as easy as is being suggested, for >Computers to usurp the current Human World Chess Champion. > >Here are some points I have collected over the years that lead me to say what I >do: > >1. Faster and faster speeds are not the B-all and End-all of chess calculation. >Chess knowledge also counts for a lot. In the beginning an increase in speed >did give a noticeable increase in results. But this was at lower ratings where >increases in speed can show good improvements. Eric Hallsworth believes the >same thing and has said so in his magazine (Selective Search) on numerous >occasions. > >2. We appear to have two main types of program, speed programs and knowledge >programs. The speed programs do not have extensive chess knowledge, therefore >fewer programming instrauctions therefore this program type can work through the >plies faster as there are not so many program instructions to execute against >each position. Knowledge programs on the other hand have relatively extensive >chess knowledge and work through the plies more slowly. They are slower for the >obvious reason that there are more program instructions to execute against each >ply. Has it ever been documented what a so-called knowledge program actually does with all of that time? Is it really spending time analyzing positional features, is it trying to identify tactical opportunities, or is it trying to selectively prune? I think people are over-eager to draw conclusions based upon node count and author descriptions. I think if you took a so-called speed program and modified it so that the count outputed to the screen was divided by 32, everyone would rave about the program's new knowledgeable style. Mine is a lot slower than it used to be, but I didn't add any knowledge to achieve this, I just changed the shape of the tree it searches. >3. The law of diminishing returns appears to have come home to roost on the >speed programs. Even with increases in processing speed the knowledge programs >are winning more often. I don't think this is supported. >4. Although mainframe programs have received the glory for apparent amazing >results the games have been carefully engineered (i.e. no prematch analysis of >previous computer games, small game sample, non-tournament time controls, and no >rematch situation) How many games was Anand allowed against the configuration of Rebel that he played against? >5. PC programs have far superior knowledge to the mainframe programs. >Unfortunately because they run on slower hardware, and don't have the financial >backing that some mainframes have it has not been possible to convert a PC >program to run on a fast mainframe. > >6. If you gave the strongest GM in the world the opportunity to prepare properly >for a tournament against the strongest computers, with tournament time controls, >sufficient games to eliminate small sample freak results, access to all previous >games played by the computer (which is exactly what current human players have) >I believe there woul dbe no contest. Computers will continue playing with the >instructions they have, they cannot change their instructions, so a mistake is >always going to be a mistake. So computers assess and mark an error and will >not play that error again, but that depends on an astute assessment algorithm >that understands where the error occurred. Programs are not sophisticated >enough to determine the correct lines to avoid when errors occur in previous >games. But that is exactly where humans have their strength. In their ability >to change their approach and learn profoundly from their mistakes. Computers >are not in that league at the present. A Gm will play a few games and determine >where the weaknesses are then play to exploit the weaknesses, i do not see >computer programs, at the present time, using that technique to build up a >porfolio, if you like, of its opponent and forming a strategy to home in and >exploit the weaknesses. I agree with the foundation of this. There are a series of domains you can focus on, when you research computer chess: 1) Individual positions. 2) Single games. 3) Arbitrarily long matches, interruptable by programmer. 4) Fixed-length matches. 5) Career. A human chess player is self-deterministic and operates in the career domain. The vast majority of research in computer chess has been focused somewhere between individual positions and single games. Automatic play on chess servers has to involve the arbitrarily long match domain, at least to a degree. Some people will point to this distinction and accused the computers of cheating or at least inferiority. This will happen more often as they do better and better in the single-game domain. But that's just where they are now. That doman is hard enough. If sufficient strength is present there, it is trivial, through use of book learner, to rig the thing up to survive in the fixed length match domain. This is simply a matter of program stability, a reasonable duration of the match, a diverse opening book, and a little luck. Winning individual games is a lot harder. There will always always always be grounds for accusation that computers are cheating. Even if you produce one that mimics a grandmaster in every respect, there will be some differences that will cause a complaint. And even if there is no difference, the humans will say that the very non-humanity of the machine is an advantage. In short, there is no way to win this, it's simply a matter of dealing with inevitible indignation. Kasparov's reaction to losing against DB was entirely predictable, I just wish I'd made these comments before that match rather than after. >7. I feel we may be reaching a plateau in the development of chess programs. It >is becoming more difficult for chess programmers to find improvements in their >programs that produce a gain more than 50-80 ELO at best. 50-80 Elo is a *lot* for one version of a mature program. It might be interesting to add up all of the advertised Elo gain between various versions of a program, then subtract out that which can be shown to be due to hardware, and see how much is left. I think 50-80 is probably way too much. But it's possible to make incremental improvements and to enjoy the larger improvements that are due to hardware, and this should go on for some time. Additionally, the Elo rating of these programs is so large already that even a little increase is interesting. bruce >8. the programmers are finding it tougher to build more strategy awareness into >the programs. It may be some years before strides are made forward in this >area. > >9. There is now a very real threat to further development of current chess >programs due to the price war currently raging in the industry which makes >return for investment very difficult to achieve. I am saddened to see this >happen. > >With these points I believe we have very real problems in developing a far >stronger program than we have at present. > >I hope we do overcome these problems but at the present time I cannot see a way >forward for large increases in strength. > >I reckon the best GMs can sleep at night they are not about to be ousted by a >computer world champion in the near future. > >Comments anyone. I'd be happy to read some. > >Keith Kitson
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.