Author: Peter McKenzie
Date: 17:02:23 05/24/00
Go up one level in this thread
Sounds like great work Ernst, I look forward to reading it. I've always wonder though if fixed depth games were the right way to go about looking for dimishing returns. To me, it seems more useful to use fixed node counts (eg. 1 program is allowed N million nodes per move, vs same program being allowed M*N million nodes per move). This would model more closely what happens when we get faster hardware. The problem with fixed depth is that the endgame is played at a much worse level than the rest of the game. Well, worse in relation to what a chess program playing under a normal time control would do. I mean, normally programs search much deeper in the endgame. Perhaps my concern is irrelevant, I'd like to hear your opinion. Sorry if you have addressed this issue already, i'm unable to read your article right now. cheers, Peter On May 24, 2000 at 15:08:15, Ernst A. Heinz wrote: >Dear Fellow Computer-Chess Enthusiasts, > >In view of the current discussion about diminishing returns in the thread >"Ply Depth in Relation to ELO again", I like to share the results of my >latest self-play experiment with you. > >The stunning outcome of the new experiment is that it shows the existence >of diminishing returns for additional search in computer chess self-play >with 95% statistical confidence, exemplified by the program "Fritz 6"! > >The title and abstract of my M.I.T. LCS Technical Report on the >experiment follow below. > >*********************************************************************** > > ``A New Self-Play Experiment in Computer Chess'' > > ABSTRACT > >This paper presents the results of a new self-play experiment in >computer chess. It is the first such experiment ever to feature search >depths beyond 9 plies and thousands of games for every single match. >Overall, we executed 17,150 self-play games (1,050--3,000 per match) >in one "calibration" match and seven "depth X+1 <=> X" handicap >matches at fixed iteration depths ranging from 5--12 plies. For >the experiment to be realistic and independently repeatable, we relied >on a state-of-the-art commercial contestant: "Fritz6", one of the >strongest modern chess programs available. The main result of our new >experiment is that it shows the existence of diminishing returns for >additional search in computer chess self-play with 95% statistical >confidence, exemplified by the program "Fritz6". The diminishing >returns manifest themselves by declining rates of won games and >reversely increasing rates of drawn games for the deeper searching >program versions. The rate of lost games, however, remains quite >steady for the whole depth range of 5--12 plies. > >*********************************************************************** > >Please find the full report in gzip'ed PostScript format at the URL ><http://supertech.lcs.mit.edu/~heinz/ps/new_exp.ps.gz>. > >Any comments welcome! > >=Ernst= > >P.S. > >Electronic preprints of my earlier publications on the relationship >between computing power and playing strength of chess programs are >available from http://supertech.lcs.mit.edu/~heinz/ and the WWW >pages of "DarkThought" at http://supertech.lcs.mit.edu/~heinz/dt/.
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.