Author: Robert Hyatt
Date: 15:09:57 12/10/97
Go up one level in this thread
On December 10, 1997 at 15:35:23, Howard Exner wrote: >On December 10, 1997 at 13:48:00, Bruce Moreland wrote: > >> >>On December 09, 1997 at 18:37:37, Don Dailey wrote: >> >>>I never seriously considered playing my program Cilkchess on the >>>internet >>>for these very reasons. It's probably a good thing if all you really >>>want is to get in some games in and find weaknesses, but I get very >>>uncomfortable with ambiguious results. >> >>This was frustrating to read, given that it was prompted by something >>that I wrote. >> >>My program has been on the net for three years, and having it there has >>helped me a lot. I wish that you'd get yours on the net as well, and >>would hope that others would also do this. >> >>I get a wide variety of opponents. Sometimes you knew exactly who you >>are playing, and get a good idea of how you stand against that account. >>Other times you don't know what bet you -- it could have been Roman >>playing on a ghost accounts, it could have been another computer plus a >>passive operator, it could have been a human assisted by a computer, or >>it could have been some fish who found a hole in your program. So you >>get a wide variety of losses to learn from. >> >>For about a month before Paris I looked at every game, every day. Since >>then I've taken a break, but I have to catch up. I have 1114 games to >>look at (at least 260 with IM's or GM's, and a couple hundred with other >>computers) , including 196 losses or draws. This will take a while, but >>I'm going to look at each of those games, starting with the losses, of >>course. I'm sure I will plug some holes in the book, discover some >>middlegame mistakes, and get some ideas for more endgame knowledge. > >Could you post a summary of what you discovered in the lost games. I >don't mean posting the pgn's but stuff like how many duplicate opening >busts where encountered. >> >>I think this is very worthwhile, considering that all I had to do to get >>this data is connect to the net and wait. >> >>How do you get games? Do you auto-test? Do you have strong players >>that play a lot of games with your program? >> >>bruce I can't speak for Bruce, but I also look at *lots* of games. Here's the sort of things I find: 1. opening busts. This generally isn't a problem due to book learning, but on occasion, someone will find some three move sequence to take crafty out of book *very* early, and then go fishing around for a way to win. Once this is found, they will do it every other game. Position learning helps, but doesn't fix everything. I don't find many of these however... 2. single tactical losses. IE a position where something suddenly goes wrong. These are often separated into two cases: (a) simply needs more speed or depth to fix; (b) knowledge could help avoid the loss. In (b) I can do something. in (a) I can either try to tinker with the search ex- tensions to make it see the problem quicker, or else ignore it, which I generally tend to do... 3. common "thread" losses. IE Crafty used to let the opponent create an outside passed pawn, then trade down because it was a pawn ahead, only to find that the resulting K + P ending was lost. I added some knowledge for outside passed pawns. Then it started screwing up by having an outside passer itself, but the opponent had a protected passer somewhere else. The outside passer then doesn't win, yet it played right into it thinking it was doing well. This is a never-ending ladder of adding knowledge to plug the holes. (3) is the case I like to look for. IE is it getting attacked and losing a lot? If so, king safety needs work. If it is getting repeatedly killed in a particular ending, that needs work. (The "mercilous" attack which opens the h-file is a classic example). But there is another big payoff besides this. The number of games, the quality of the opposition, generally means that I don't participate anywhere with a "buggy" program. Perhaps, as in Paris, it is poorly tuned due to mistakes on my part. But the program *never* crashes, hangs or burps during a game in a serious event. Because it plays *so* many games it is unlikely that any section of the code has been not executed.
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.