|
David E. Welsh wonders out loud about the computer
chess programs of the future. Now (2007) is the future he thought about then (1984). Mr. Welsh was convinced that the
programs we have today would be more interesting and more advanced. Would he have been disappointed?
p.ix"The
programs of the future will not only be stronger than those of today, but more interesting as concepts advance. Their development
is going to be an interesting story, and it will be fascinating to watch it unfold."
A great savings in search efficiency is achieved if our ordering
of the moves searched is improved, according to Welsh. It would seem that this objective should be a top concern of a chess
program developer. What calculation can we do to improve the ordering of the moves, without falling into the recursive trap
of further search?
p.86"The savings achieved by the alpha-beta [search] algorithm are
enormous if the ordering of the moves is good."
Once again, we have an author who speculates about what it would
take to make a computer chess program more 'human' like in its playing style.
p.90"The author's personal (and debatable)
opinion is that the real criterion for judging whether a program may be said to be playing chess in the same manner as humans
do, is whether it follows processes equivalent to those used by human players... The essential characteristics of human chess
play involve comprehending the game as a continuous sequence of real and possible positions interrelated by real and possible
moves. Human chess players continually examine positions to try to ascertain the essential characteristics, which are used
to set goals for their play, and in trying to achieve these goals, formulate plans... No program has yet [1984] demonstrated
much more than rudimentary capabilities in these areas, and it is fair to say that, according to this viewpoint, none of them
is yet playing chess in quite the same manner as a human does."
What would happen if we began to teach computer programs to recognize
goals and objectives? For starters, you would have a different kind of program:
p.92"computer chess programs - at least those yet developed - know
little about why they decided on their last move; nor do they (with few and minor exceptions) form plans, set goals, or credit
the opponent with forming plans and setting goals. Every position is in effect a new game, and the requirement is to maximize
a number representing a position score, not to achieve an objective dictated by the position.
The [computer] program may not know what an objective
is, let alone how to achieve one... As a result, computer play is opportunistic, inconsistent, planless, and usually materialistic
to the nth degree."
The older programs had trouble determining positional compensation,
and therefore material was made the dominant term in the evaluation function:
p.93"The reason for this emphasis on
material is that computer programs cannot be trusted with determining whether a positional consideration (e.g. an attack)
is worth a material sacrifice... In general, however, chess programs cannot judge whether positional factors compensate for
material disadvantage, and so material factors and positional factors are scored separately, with material deliberately made
dominant to prevent sacrifices played at random."
Welsh speculates that programs developed using newer techniques
might need to go through a developmental period where they play sub-par chess, compared to existing programs.
p.95,106-107"It is very unlikely that
any program that attempts to follow the human approach to chess-playing will be able to cope with the best programs of the
present type for quite some time after its introduction, if ever... early efforts with the new concepts that are needed will
probably involve a significant decline in playing strength from present fixed-heuristic levels."
Welsh describes the planless play of existing
programs, which often play in a haphazard style that is inconsistent from move to move. When tactical moves are not playable,
the machine simply cannot understand how to build positional pressure.
p.96"During the middlegame, computers
play opportunistically and without any plan other than to maximize the material and positional score at each turn. Often this
involves shifting from one expected sequence of moves to something completely different at the next turn, without any reason
discernible to the human mind. The total planlessness and lack of any aim or goal in computer play means that even the best
machines are surprisingly vulnerable to long-range strategy... Most programs have very little ability to conduct a positional
attack when no obvious weak point exists... As compensation for their lack of long-range planning and positional judgment,
chess programs - even the commercial microcomputers - are sharp tactically."
| 
