Author: Andrew Dados
Date: 11:37:13 09/14/00
Go up one level in this thread
On September 14, 2000 at 14:14:31, Andrew Dados wrote: I also noticed one interesting thing about your data: % of fail low is very constant through all depths - meaning you probably finish search for current move if time runs out. That column contradicts % of fail high column. However that all maybe 'build' in the game itself. Player can not change 'ideal game score' (as in TBs) with his move - he can only retain it or blunder to worse score. That itself can be a 'proof' for diminishing returns effect - since there are several equally 'best' moves in most positions.... -Andrew- >On September 14, 2000 at 13:42:49, Ed Schröder wrote: > >>In the upcoming Rebel Century 3.0 I have implemented a little >>statistic routine that reveals something about the nature of SEARCH >>that could be important for the future of computer chess in the sense >>that it says "something" one may expect in the near future because >>of faster and faster PC's. >> >>Research on this issue have already been done by Bob and Ernst and >>it has made me curious so I have spend a little time on it. The >>statistic shows 2 things: >> >>a) number of fail-low's for each depth; >>b) number of "changed moves" for each depth. >> >>(a) is not so important as often fail-low's do not mean anything but I >>wanted to know anyway. >> >>(b) is extremely important as it shows for each depth how many times >>Rebel changed its mind. As you can see in the below statistic the % >>diminish and diminish the deeper Rebel goes. >> >>How to read the overview: >>- first column: iteration depth; >>- second column: number of times the depth was reached; >>- third column: number of fail-low's; >>- fourth column: percentage of fail-low's; >>- Fifth column: number of changed moves; >>- Last column: percentage of changed moves; >> >>SEARCH OVERVIEW >>=============== >> >>1 4726 0 ( 0) 0 ( 0) >>2 4726 1000 (21) 1889 (39) >>3 4726 495 (10) 1468 (31) >>4 4719 209 ( 4) 1085 (22) >>5 4719 218 ( 4) 1222 (25) >>6 4699 191 ( 4) 1139 (24) >>7 4655 141 ( 3) 948 (20) >>8 4572 109 ( 2) 837 (18) >>9 4457 79 ( 1) 777 (17) >>10 3998 86 ( 2) 644 (16) >>11 3015 64 ( 2) 374 (12) >>12 1904 55 ( 2) 204 (10) >>13 1093 37 ( 3) 77 ( 7) >>14 584 22 ( 3) 35 ( 5) >>15 356 15 ( 4) 22 ( 6) >>16 230 7 ( 3) 6 ( 2) >>17 157 6 ( 3) 2 ( 1) >>18 123 6 ( 4) 3 ( 2) >>19 88 0 ( 0) 0 ( 0) >>20 67 0 ( 0) 1 ( 1) >>21 55 0 ( 0) 1 ( 1) >>22 54 0 ( 0) 1 ( 1) >>23 50 0 ( 0) 0 ( 0) >>24 47 0 ( 0) 0 ( 0) >>25 40 0 ( 0) 0 ( 0) >>26 30 0 ( 0) 0 ( 0) >>27 28 0 ( 0) 0 ( 0) >>28 22 0 ( 0) 0 ( 0) >>29 19 0 ( 0) 0 ( 0) >>30 14 0 ( 0) 0 ( 0) >>31 14 0 ( 0) 0 ( 0) >>32 13 0 ( 0) 0 ( 0) >>33 12 0 ( 0) 0 ( 0) >>34 12 0 ( 0) 0 ( 0) >>35 10 0 ( 0) 0 ( 0) >>36 10 0 ( 0) 0 ( 0) >>37 10 0 ( 0) 0 ( 0) >>38 10 0 ( 0) 0 ( 0) >>39 10 0 ( 0) 0 ( 0) >>40 8 0 ( 0) 0 ( 0) >>41 8 0 ( 0) 0 ( 0) >>42 5 0 ( 0) 0 ( 0) >>43 4 0 ( 0) 0 ( 0) >>44 3 0 ( 0) 0 ( 0) >>45 3 0 ( 0) 0 ( 0) >>46 2 0 ( 0) 0 ( 0) >>47 2 0 ( 0) 0 ( 0) >>48 2 0 ( 0) 0 ( 0) >>49 2 0 ( 0) 0 ( 0) >>50 2 0 ( 0) 0 ( 0) >>51 2 0 ( 0) 0 ( 0) >>52 2 0 ( 0) 0 ( 0) >>53 2 0 ( 0) 0 ( 0) >>54 2 0 ( 0) 0 ( 0) >>55 2 0 ( 0) 0 ( 0) >>56 2 0 ( 0) 0 ( 0) >>57 2 0 ( 0) 0 ( 0) >>58 2 0 ( 0) 0 ( 0) >>59 2 0 ( 0) 0 ( 0) >>60 2 0 ( 0) 0 ( 0) >> >>After iteration 19 one hardly sees any changes. One might wonder >>if a doubling in speed is still good for 50-70 elo (as often said >>here). >> >>One thing that should be added to the overview is a division in >>middle-game and end-game (I think). >> >>The overview was created by playing a 40/40 auto232 match (about >>50 games). Results are automatically kept so that a next auto232 >>match the statistic is automatically updated. >> >>Also if you are playing your normal (manual) games or analyze positions >>the statistic is maintained. It would be nice to see how the statistic >>would look like after say 100,000 moves. >> >>Ed > >'the depth was reached' <> 'the search for that depth was finished'. >Thus there is some chance that your 'diminishing return effect' data can be >misleading. > >Could you possibly modify your code not to count unfinished plys? That would >give much consistent data imo. (Note that given about 1/2 time is spent on best >move (PV) search, and considering random time of exiting search, data for >unfinished plys could be around 2 times higher if last ply searches were >actually completed). > >-Andrew-
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.