Author: Gordon Rattray
Date: 07:50:52 04/18/01
Go up one level in this thread
On April 18, 2001 at 04:47:04, Ralf Elvsén wrote: >On April 17, 2001 at 17:45:55, Gordon Rattray wrote: > > >>I find this interesting since I have recently participated in a postal >>tournament where the use of computers was expected by all competitors - so no >>immoral cheating was going on. I wanted to investigate how I could be use my 2 >>PCs and numerous chess programs to find best moves given a significant length of >>time. I experimented with the following approach. >> >>I had one program running continuously on one PC for the duration of the >>analysis period. The idea being that some moves may require a certain depth to >>be reached and it may be unwise to cut all programs short. >> >>My second PC was used more in a swapping fashion with multiple programs being >>given a time slot one after another. I record details such as score, principle >>variation, etc. for each one. >> >>If all programs agree, including the one not being swapped out, I play that >>move. My hope is that if the relatively deep search in combination with using a >>variety of programs all agree, the chances of missing a better move have been >>minimised. >> >>If there is disagreement between the programs, I draw up a list of candidates >>based on all the suggestions. I then for each program show it the candidates >>that it didn't agree with and I record how much of a difference there was >>compared with the program's initial choice. At this point some programs "admit" >>the candidate is better by showing a better score. For those who show a big >>drop, I play out more of the principle variation in hope that one of the >>programs in question will prove itself correct or wrong. >> >>The above still leaves a lot of cases where there remains disagreement. I >>generally find myself making up variation trees with each program getting a >>chance to evaluate the leaf positions. So no great surprises here, as I often >>have to use my own judgement, but this approach does give me a lot of >>information to guide me. Deciding on a when to swap programs is naturally >>tricky. I give a preference to the unswapped program unless any of the other >>programs find a "better" move that I failed to disprove. >> >>Of course, some best moves will still be missed. If only a guaranteed approach >>existed... ;-) This is my "first draft" approach and I'm looking for ways to >>improve it, so any suggestions/comments will be appreciated. >> >>Gordon > >Nice to see someone taking a systematic approach to this. You mentioned >you have to use your own judgement, so the better operator the better¨ >result? Would it be possible to implement this as a mindless algoritm >so a patzer would get the same result? I guess not since you can't compare >evaluation scores between programs. Automating this further would be difficult. As you mention, evaluation scores do have different meanings between programs. It's also important to consider the time spent and the depth reached. I often look at last several analysis lines to see if the move has been changing. If it has been changing I may decide to allow it more time until it becomes more stable - but of course it can always change at anytime and the analysis time is limited. > >When using your judgement I assume it is important to have an idea >about the strengths/weaknesses of the various programs. E.g. in >some endings I wouldn't trust some programs and when e.g. launching >pawn storms I wouldn't trust others. Yes. When disagreements do occur, I have to use my knowledge of which programs are best for certain positions, to some extent. Unfortunately this knowledge can never be exact so I try to get the programs to prove themselves by analysing each others suggestions, possibly playing down the principle variations as far as time will allow. Sometimes it works and sometimes it leads to more disagreement. Also, it has to be remembered that many positions don't have a single best move as such and that there are multiple good moves. I'm thinking of adding other techniques to my approach. Maybe for certain positions I should allow it to be played out to an end, using different programs for both sides, and see which move has the highest win ratio?! This might highlight which program can best play the position. Also, the Chessbase interface has a correspondence analysis option, so maybe this can automate some of the effort?! It's a tricky problem trying to make the most of the time. Gordon > >Ralf
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.