Author: Dann Corbit
Date: 21:25:32 10/01/99
Go up one level in this thread
On October 01, 1999 at 22:17:33, Ratko V Tomic wrote: >> But if you do something really strange like what I did, then >> all bets are off. I had to modify Crafty to make it work. >> >> It also plays something super brilliant once in a while, so there >> are no guarantees you will beat it! > >That's some number of patterns, with 2 million of games there must be >at least 150-200 millions of positions. There is probably every strategic >and tactical maneuver of practical use in there, as well as much of >the interesting endgame positions. If only one could use more than just >exact match for the openings, it could make some truly humanlike play. It does not necessarily have to use exact matchups. For instance, we are at some position 'P' and you let your move generator spew out moves for P. Now, ten of them are in the database. So you look at them one at a time. Three of them look pretty good, with a won/loss/draw figure showing good advantage. So you just take the best one of them and do it. Your opponent responds (perhaps in a way you did not expect) putting you at a new position Q. You just repeat what you just finished doing until you have no more hits in the database. In that way, you will always try what has been successful. >The problem with current programs is that they don't make use of openings >the way humans do, i.e. by keeping in mind typical patterns or themes >of play following a given opening line. Humans get this by examining >complete games of a given opening, so they can see the purpose and potential >of the achieved configuration. Some sort of dynamical evaluation function >(which might work only at the root level), tuned to the sample games from a >given opening (e.g. to value more certain piece/pawn activity, or control of >given squares, or value of pieces and an "urge" to move given piece), should be >able to pick up some of that pattern from the remaining parts of the games. (If >I ever get my program beyond the planning and gathering phase...) The big problem with the approach listed above is that what has worked well may have a serious flaw upon close examination. It seems that no human has ever seen the forced skewer that can be formed 6 moves down the road against you if you do it. A computer can find that move and make you pay for it. Hence the need for 'learning' files like Crafty uses. There are other ways that the opening book could be improved. ;-)
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.