Computer Chess Club Archives


Search

Terms

Messages

Subject: Engine ratings vs humans: experiment design

Author: Edward Seid

Date: 06:40:03 12/25/01


Hello,

In mid-January, I will start the Man vs Machine Rating Project on ICC.
Concentrating on Winboard engines, the project's purpose is to formulate an
engine RATING LIST BASED ON PLAY AGAINST HUMANS ONLY.  I'm trying to determine
what things to keep constant in order to make it fair for all engines tested.

The basic project design is as follows:

- each engine will have its blitz rating initialized to an estimated strength
- all games will be 2 12 rated, against humans with an established blitz rating
- human opponents are limited to +/- 200 points from engine's current rating
- maximum of 4 consecutive games against same opponent
- hardware platform to remain constant (current configuration is 1.0 GHZ AMD
Athlon, 768 MB RAM, Windows 2000 Pro, Winboard 4.2.5, cable modem internet
connection. subject to change before start of experiment)
- each engine will be allowed to run continuously for 7 days, resulting in
500-700 rated games
- games with no result or <5 moves will be discarded. the remaining games will
be collected and made publicly available
- engine parameters to be kept constant (hashtable size, use of opening book,
ponder mode, use of EGTBs). [opening book set to largest available, ponder ON if
allowed]
- any adjourned games will be discarded before testing the next engine

My questions are the following:

- time control: 5 12 also being considered.  pros/cons?
- hardware: also have weaker PII/233 with 128 MB RAM running Windows 98
available.  pros/cons to running experiment with weaker hardware?
- hashtable size: deciding between 64, 96 or 128 MB
- hashtable allocation: if configurable, how to allocate total hash between
hashsize, q_hashsize, pawn_hashsize?
- EGTBs: currently deciding between 3/4 complete or 3/4/5 complete.  pros/cons?
also, how many MB to allow for EGTB (not included in total hashsize)
- resign value: trying to decide between -7 and -9. comments?

Any comments/suggestions will be greatly appreciated.  Thank you in advance.

ed seid

PS  If you're a Winboard engine author and would like for me to run your engine
in this experiment, please send me an email at manvsmachine@winboard.info and I
will add you to the priority list.



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.