Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Handicapping Chess Engines

Author: Robert Hyatt

Date: 12:40:33 10/26/01

Go up one level in this thread


On October 25, 2001 at 16:20:42, John Merlino wrote:

>On October 25, 2001 at 12:02:29, Robert Hyatt wrote:
>
>>On October 25, 2001 at 06:37:00, Mike Hood wrote:
>>
>>>There have been a lot of posts (over the years) about how to make engines
>>>weaker. This is an interesting question that I'd like to revive. I mean, it's
>>>not just about getting to the top of the SSDF list. Very few people who buy a
>>>current chess program at its full unleashed power stand a chance of even drawing
>>>against it. It's a matter of playing enjoyable (and instructive) chess in your
>>>spare time.
>>>
>>>I'm disappointed with most of the handicap levels in chess programs. Everything
>>>from "Drunken Assassin" to "Paranoid Scaredycat" delivers poor chess, making
>>>stupid blunders that anyone with an ELO rating over 1000 can take advantage of.
>>>The only successful handicapping that I've seen is limiting the ply depth of the
>>>search. Limiting the ply search of an engine to 4 (or maybe 6) ply leads to the
>>>engine playing solid but beatable chess. Maybe I'm just speaking from my own
>>>limited perspective as a 1550 player, but I have the impression that the
>>>blunders made by a plydepth-limited engine are very "human".
>>
>>I don't think this works.  I ran some tests on one of the chess servers once
>>with a very limited search depth and still saw ratings of over 2200 at times.
>>
>>Bruce ran a version of Ferret on ICC with a time limit of milliseconds per
>>move and it too was in the 2200 range if I recall correctly.
>>
>>The problem is that against weaker players, even shallow searches see through
>>their simple tactics, and you _still_ have the full positional evaluation of
>>the program, with weak squares, pawn structure, king safety, etc. to guide the
>>program.
>>
>>If you reduce the search depth, you only reduce tactics.  You won't find a
>>1200 player that understands much about pawn majorities, weak pawns, king
>>safety, etc.  You end up with a program that is tactically weaker, but still
>>positionally very strong.  And (IMHO) it just doesn't "feel right".  IE weak
>>players will let me wreck their pawn structure where a shallow search program
>>will not.
>
>We have found this also to be the case with Chessmaster. There are a few
>personalities that are identical to the Chessmaster personality, but have a VERY
>limited search depth. Vlad has a search depth of 3, and his rating came out to
>approximately 1850. Max has a depth of 2, with a rating of around 1600. And
>Lacey (the bane of all existence) has a depth of 1, with a rating of around
>1300. So, I would estimate that a search depth of 5 would be approximately in
>the 2200 range that you mention above.
>
>Does that come close to matching your findings?
>
>jm


I haven't tried fixed depth very much.  Because it is _way_ too "uneven"
overall.  IE N plies in the middlegame might play at rating X, while
N plies in the endgame plays at rating X-N because in the endgame X plies
go by much quicker.

If I had a lot of time, I would try to set up a few ICC accounts that played
at specific levels.  And tune them to do so.  However, I think it would be a
non-trivial task, particularly to reach lower levels like 1300 - 1600 or so.
And then there is the hardware issue that would require some sort of quick
benchmark to establish some speed reference point.



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.