Author: James T. Walker
Date: 20:04:14 12/09/00
Go up one level in this thread
On December 09, 2000 at 22:27:31, Jeremiah Penery wrote: >On December 09, 2000 at 20:32:58, William Penn wrote: > >>CM8000 position evaluations, +- Scores >> >>The position on the chessboard can be scored as either + or - depending on which >>side has the advantage. + values favor white, and the larger the value the more >>favorable for white. - values favor black, and the more negative the number the >>more favorable for black. This is an absolute protocol to express the current >>status of the game, and has nothing to do with which side is on the move. Any >>variations from this widely understood protocol should be carefully explained, >>or confusion is likely to result. >> >>I have observed a related problem with CM8000 more than once. Simply, you can >>ask a custom personality to analyze a position for you, and it will yield >>certain results. Then you should be able to confirm those results by making >>each possible move from the position, rather than letting the software do the >>iteration of possible moves. That should usually give you approximately the >>same results at one less ply. The problem with CM8000 is that the signs of the >>resulting scores are apparently wrong, sometimes. Instead of + they are -, and >>vice versa. CM6000 got it right. CM8000 gets it wrong. Here's an example... > ><snip> > >> after 10.e3 after 10.e3 >>CM6666 Deep (CM6K) CM8666 Deep (CM8K) >>Depth Score Moves Depth Score Moves >>01/09 -0.52 Ne4 9 0.54 Qc7 >>01/10 -0.41 Ne4 10 0.33 Qc7 >>01/10 -0.47 Rc8 10 0.42 Qa5 >>01/11 -0.39 Rc8 10 0.46 c5 >>01/12 -0.29 Rc8 10 0.50 Rc8 >>01/12 -0.36 Qc7 11 0.44 Rc8 >>01/12 -0.43 Ne4 12 0.39 Rc8 >>01/12 -0.46 b5 12 0.46 Ne4 >>01/13 -0.29 b5 13 0.42 Ne4 >> >>We see (above) that there are many possible moves with similar scores. Also we >>note that the CM6000 scores are all negative, but the CM8000 scores are all >>positive. In the series of moves at 01/10 and at 01/12 depths with CM6000, the >>changes are towards more negative scores. That seems proper because we are >>searching for the best black move. (better black positions are more negative) >>In contrast the trend of scores with CM8000 in the series at 10 and 12 depths is >>towards more positive values. That appears to be in the wrong direction when >>trying to find the best black move. The reason is probably that CM8000 fails to >>put the - sign on the moves here after 10.e3. They should probably all be >>negative values, not positive values, analogous to the results with CM6000. > >I'm not entirely sure what to think here. The depths you show (all being 1/xx) >really look strange. At depth 2, it should show 2/xx or something similar. I >can't tell if this is real search output, or a 1-ply search output, or something >more like a root movelist. > >However, I'll venture the guess that CM8000 is showing the evaluation from the >point of view of the side to play, while CM6000 showed it from white's >perspective always. Another possibility is that the evaluation is just a lot >different in this position. CM8000 is a different engine version than 6000 was, >so that could explain part (if not all) of the difference. No,both programs show +x.xx for white winning and -x.xx score for black winning. The only thing I see unusual about the above is that CM8K thinks that Re1 is the best move but when you make the move e3 the score increases for white indicating that it's a better move than Re1. So why didn't it play e3 instead of Re1? As far as the changes go, there is no way to make it get better at deeper depths just because that's what you are trying to do. Whether you like it or not at deeper depths you may find that your score is just getting worse and there is nothing you can do about it. The deeper scores are just based on more information and it may be good info or bad info. I don't have the settings for these personalities but the plain CM8K likes Re1 with a score around 0.03 but after I play e3 the score goes up to about 0.20 which as I said indicates it likes e3 better so why does it want to play Re1? I suspect only the programmer can really tell you the answer to this. Jim
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.