Computer Chess Club Archives


Search

Terms

Messages

Subject: Worthwhile Research on Computer vs Human Styles

Author: Bob Durrett

Date: 17:25:42 11/11/03


Almost every day, some one here posts a position taken from a GM game, or from a
historically significant game of past human chess giants, and shows that chess
engines have a hard time with that particular position.

The idea seems to be that collections of such positions serve well as test
positions for programmers who are trying to improve their programs.

Generally, if a position is found where no engine would play the move the human
played, one might suspect that the human simply erred.  However, this can be
analyzed and the truth be found out.  Only test positions with moves fully
verified are suitable.

It may be, however, that a strong human move is unsuitable for a chess engine.
Training one's engine to always play moves good for humans might make the play
of the engines more "human-like" and eventually it might become impossible to
distinguish the play of the engines and humans.  Someday.

However, engine versus engine competitions may require a different kind of
engine to get to the top of the rating list.  It may be that engines optimized
to play "human-like" would perform badly in SSDF competitions, for example.

Perhaps it would be worthwhile to create an engine which plays EXACTLY like a
human.  It would certainly cool the heels of those sick moralists!  But, beyond
that, it might help to advance the state of the art in engine design.  As a
trivial example, since humans are supposed to be better in long-range planning,
then a truly human-like engine would have to be equally capable at such
planning.

One approach to make engines play good "engine-like" games would be to work on
reducing the time required to find the best move.  For example, if it took 15
minutes for a specific chess engine to find it's best "engine-like" move, then
the programmer might wish to strive to improve the software so that the move
would be found in half the time.

Finding good "engine-like" moves is really easy, albeit time-consuming.  Simply
let a dozen different top engines evaluate a position for a long time and then
let each engine furthur evaluate the choices identified.  This process could be
continued to create a tree of lines which are long enough to preclude "horizon
effect" errors.  Etceteras.

Once such a "hard-to-find" move is identified, then the programmer could try to
find ways to make his/her software find the move quickly.

Such a procedure might eventually produce engines which quickly generate very
strong "engine-like" moves which may or may not be at all "human-like."  Then it
might be trivial to win top spot on the SSDF list.  Who cares about humans?
They're just carbon-based!  Ugh!

: )

Bob D.




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.