Author: Dann Corbit
Date: 18:10:37 07/23/02
Go up one level in this thread
On July 23, 2002 at 21:00:49, Pham Hong Nguyen wrote: >On July 23, 2002 at 18:28:38, Dann Corbit wrote: > >>On July 23, 2002 at 17:48:09, Shane Hudson wrote: >> >>>On July 23, 2002 at 03:11:17, Dann Corbit wrote: >>> >>>>On July 23, 2002 at 03:04:39, Pham Hong Nguyen wrote: >>>> >>>>>Hey Dann, I did not say about your test set - yes, all games of that test set >>>>>were artificially created to illustrate how important underpromotions are. >>>>>However, I am mentioning about _real_ games - other story. >>>> >>>>Almost all of those positions I posted are from *REAL GAMES*. Only a very tiny >>>>fraction are made up positions. >>>> >>>>>If you or someone could (or I will do when I have spare time), do a small >>>>>research about underpromotions in real games (of any database): count and report >>>>>total number of underpromotions / total number of promotions. The statistics >>>>>will help us on decision of design. >>>> >>>>The amount of useful underpromotions in real games will definitely be higher >>>>than you think. If it is as small as one in a million games where >>>>underpromotion provides benefit, I will be utterly astonished. >>> >>>I investigated this once for the purpose of improving material and >>>position search times in my database app Scid. >>> >>>Here are stats on the number of games containing a promotion to each >>>type of piece in a database of 594,803 mostly master-level games: >>> >>>Piece Games Freq per 1000 games >>>------------------------------------- >>>Any 24747 41.61 >>> Q 24083 40.50 >>> N 506 0.85 >>> R 227 0.38 >>> B 72 0.12 >>>------------------------------------- >>> >>>Whether many of those promotions to Rook or Bishop were actually >>>useful (superior to a Queen promotion) is anyone's guess. I suspect >>>most occurred in situations where the piece will immediately get >>>taken anyway. >> >>In a database of 1.7 million games with players of 2000 ELO+, I found 3056 >>underpromotions. > >You all have excellent data. Dann, could you continue to check Shane Hudson' >idea: if those underpromotions are really necessary (if not, lose or draw) or >just because they will be taken on next move? I did a very quick check (shallow mate search) to produce the test set of underpromotion EPD positions. The link is found upstream. Now, I think a very large number of these sorts of things will also be overlooked if in the search it sees and avoids them, compared to fails to see them. They might never get played, but if not seen, the path may not be optimal. The test set I created has the following: The named move *must* be played for optimal play. I did not check to see if all other moves do worse, as far as losing or drawing. I did not examine any of the cases where underpromotion merely leads to an advantage, since those will be problematic to know for sure if it really is best. So the test set I did provide will test to see if you find the best move. I will be curious to know what an engine that cannot underpromote thinks of the positions. Crafty rushes through them with 100% correct in a flash.
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.