Author: Uri Blass
Date: 23:01:37 05/28/02
Go up one level in this thread
On May 29, 2002 at 01:27:44, Albert Silver wrote: >On May 28, 2002 at 14:18:04, Uri Blass wrote: > >>On May 28, 2002 at 13:17:57, Albert Silver wrote: >> >>>On May 27, 2002 at 14:02:47, Uri Blass wrote: >>> >>> >>>I don't understand the purpose of changing the first move because of a loss. >>>This implies that the first move was the cause of the loss, when it could easily >>>(probably) been due to an error much later. If I play a wrong move in an Evans >>>Gambit and lose because I didn't see far enough, does that mean the fault was >>>1.e4? >>> >>>If the plan is to learn from previous moves, I would imagine that if a time >>>control and the conditions (memory + CPU, etc.) is the same then it should play >>>the same first move instantly, presuming that identical conditions will produce >>>identical result, and then begin calculating again. The actual learning, to >>>rpevent repetition, should logically appear just before the move where the >>>evaluation dropped. Perhaps the program could be requested to begin studying the >>>first two moves prior to the drop in greater depth. Example, if the eval dropped >>>at move 6, then begin calculating at move 4 the same move, checking the log for >>>the last depth achieved, and requesting the program go one ply deeper to see if >>>it can detect the error. If no changes are found, then when it reaches the move >>>prior to the eval drop, it will of course exclude the losing move though >>>retaining the evaluation. Of course the request must respect the time control >>>algorithms. Finally, the reason the first move alone should be played instantly, >>>and not all the first moves before the perceived error, is that once it plays >>>one move instantly, the time left changes, and this means that the next move may >>>change now that it has more time to spend on its remaining moves. >>> >>>I should point out that I have no real knowledge of how learning is done >>>nowadays, so it is essentially speculation. >>> >>> Albert >> >>The possible problem that I can imagine is that the program gets a position that >>it does not understand and the real error is not 2 or 4 plies before the eval >>drop but 10 plies before the eval drop. >> >>Uri > >Sure, but you're not going to go so far as to blame 1.e4 for the loss are you? >What happens if it loses a game with all the tried moves. Does the program >retire and take up golf? :-) I said that the program may remember only the last 20 games with the same color for learning because after more games it is probably going to play a different opponent. > >Seriously though, I understand the fear, but if you begin to simply make the cut >too early for no good reason, you will be excluding many lines that are not only >legitimate, but good. > >I agree with Mogens that the last branch before the eval drop is logical, but it >is because I also agree with your fear, that the error may likely have come >earlier, that I proposed beginning deeper (deeper than the logs show was >analyzed) searches 2 plies before the detected eval drop. Still, simply forcing >a chnage 2 plies before with no better justification than such a fear would not >make sense. > > Albert I agree that there are better ways of learning but my point is that I believe that the Fritz book is over estimated and I think that even the idea that I suggested will not cost a program more than 100 elo. If you do the same but only let changes that do the evaluation at most 0.2 or 0.3 pawns weaker so you will never let things like 1.e4 f6 then I believe that the rating advantage of book relative to only positional learning may be even smaller(again you may forget every game that is not in the last x games when x is big enough and you may learn to repeat wins so if you find that the program wins often with 1.h3 you may repeat this move). I suspect that you may also do better by not forgetting everything. If you find that you won 10 games with 1.h3 and lost one game you may remember that statistics and use it. It means that after you forget the game that you lost by 1.h3 you may try 1.h3 again because you are probably playing against a different opponent and the statistics suggest that it is a good move for the program. Uri
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.