Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.1 6/24/83; site eosp1.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!allegra!princeton!eosp1!robison
From: robison@eosp1.UUCP (Tobias D. Robison)
Newsgroups: net.chess
Subject: What makes a strong computer chess program?
Message-ID: <988@eosp1.UUCP>
Date: Tue, 3-Jul-84 18:16:34 EDT
Article-I.D.: eosp1.988
Posted: Tue Jul  3 18:16:34 1984
Date-Received: Wed, 4-Jul-84 03:56:39 EDT
Organization: Exxon Office Systems, Princeton, NJ
Lines: 66

References:

Some of us are currently playing correspondence chess with the
"Phoenix" computer program, and some of us have other experience
playing against computers, or interest in the subject.  I wonder
if some of us would enjoy discussing the question of what a really
strong chess computer program will be like?

There was a chess match about 2 years ago between 4 candidate masters
and 4 of the best computer programs.  There was much hope for the
programs, but they only scored about 6 out of 16 points.
The match was reported in Chess Life/Review, along with speculation
that a stronger group of chess players was at last starting to
probe for weaknesses in the computer programs, with inevitable results.

Current chess programs have opening libraries, and otherwise
analyze primarily by brute force.  I believe that there is also a
tendency for programs to extra-analyze certain capture situations
and other concerns, such as passed pawns, that can cause the
"horizon effect".  ("Horizon effect" refers to a situation where the
computer ignores, say, the opponent's threat to queen a pawn because
it can postpone the threat just one move beyond the depth that it
analyzes.)

Here are some of my feelings about what strong chess programs need to
do:

(1) They do need good opening book lines, but care is necessary to
limit the book to lines that the computer feels comfortable in.
(Example -- Phoenix is playing the French defense, which leads to
cramped positions that are hard for a computer to evaluate; I don't
a computer should play 1. ... P-K3.)

(2) Computers should never be ready to repeat a game exactly in
tournament play.  In tournamement play, any good master is ready to
replay, from memory, a loss by the computer in a previous round.
(One player actually did this in the match described above.)
Correspondence players can exchange games easily.  Therefore a
chess program playing correspondence should be careful to avoid
repetition (or at least make that unlikely) in simultaneous or
sequential correspondence games.

(3) Because (as far as I have seen) computers do better in open
positions, they should have a strong preference for opening up
positions, even to the extent of making some speculative sacrifices
of small amounts of material or positional values.  Therefore
if a computer does give extra analysis to exchange possibilities,
it should give the same extra analysis to possible sacrifices against
pawn chains.

(4) Chess programs should vary the depth that they analyze positions.
It is possible to get a feeling for awhile about how deeply a
computer is routinely analyzing.  It is much harder to do this for
humans.  It may be that the best programs already do this, but it seems
to me that the programs can safely vary their rules for deciding
how to spend time in a position so that it is much harder for a
human opponent to feel safe about a particular horizon.
In particular, chess programs should be able to do something that
humans do fairly easily -- analyze long forcing and semi-forcing lines.


If I receive private mail on this subject, I will summarize for the
net.
					- Toby Robison (not Robinson!)
					allegra!eosp1!robison
					decvax!ittvax!eosp1!robison