Here we have in a very condensed form what is likely to be the
strongest argument against the success of the proposed heuristic. I should make a special note to myself to address
these concepts, because they are certain to be raised by those who do not see the concept presented in this paper as having
promise.
Powerful machines in 1985 were expensive to operate.
viii"The computer on which CRAY BLITZ runs is billed at $50,000 per hour! ...Perhaps the future
even holds a selective search that really and reliably works - in which case we will then have a computer World Chess Champion."
Previous attempts to add chess knowledge to the evaluation functions
of computer chess programs have failed because they have reduced the search depth and have otherwise weakened the machine's
ability to play a strong tactical game of chess. The gains in positional play did not offset the losses in tactical search.
In other words, the machines lost more games with the chess knowledge added. Perhaps they did not implement the chess
knowledge in the most efficient way.
p.42"Experience has repeatedly shown that attempting to give the
evaluation function of a full-width program much 'chess knowledge' is a difficult and dangerous thing to do... real-world
evaluator functions do not pretend to evaluate positions in enough detail to assign really accurate scores. The objective
is to give the program just enough information so that it can use its tree search to discover the effects of complicated dynamic
factors which the evaluator can't assess, such as the results of combinations."
Baczynskyj laments the fact that he cannot put together a concept which can be used by a computer
program to determine when sacrificing a pawn in a chess game is worthwhile.
p.102"I tried to establish some quantifiable guidelines for judging
when an attack would be worth a Pawn. Well, if such laws do exist at all, they eluded my attempt to present them in computer-usable
form."
Again, attempts to teach the computer program how to evaluate the relative strengths of minor pieces
based on subtle chess knowledge have failed due to the unintentional reduction of search depth without a corresponding positional
gain.
p.103"One cherished theme of my own chess play is the value of a minor piece (Bishop or Knight)
in a specific position. Computer programs use a fixed value for each piece. In practice, though, Bishops may be 'good' and
Knights 'bad' (or vice versa) depending on their location and the associated Pawn structure. My initial suggestions on implementing
variable values for the minor pieces were coded by Kathe, but proved to be so time-consuming that the tactical search was
crippled. What good will it do a program to have a sophisticated apprehension of the value of its Bishop if it loses the piece
because it 'sees' ahead one less ply [of search depth]?
This dilemma is the key limitation on the amount of 'chess knowledge' that a
microcomputer program can have. A more sophisticated evaluation function is also more cumbersome. Tactical search - the 'I
move here, he moves there, I move here and zap him' type of thinking - is the backbone of all chess play, human or machine.
'Stupid and fast' has it over 'smart and slow': this principle seems to be confirmed by the results of computer chess tournaments.
'Smart and fast' is the ideal, but probably won't be achieved in microcomputers without a hardware breakthrough... maybe,
just maybe, the whole effort to endow machines with something resembling a human chessplayer's intuition is doomed to failure"
Here we have the interesting idea that progress might be made by steering the search along the lines
that are most promising from a chess sense point of view. I strongly agree, and this concept alone can defeat the 'combinational
explosion' of the search tree by focusing the attention of the machine on the lines of play that are promising. There is only
one problem - how exactly are we to do this? The proposed heuristic offers a method for doing this, when combined
with selective search, will keep the machine focused on the most promising lines of play.
p.109-110"A key breakthrough still has to be made before chess programs
can break into the top circle... Another possible development could be some conceptual rephrasing of the minimax search, so
that a program does not waste the major portion of its time on variations that don't make chess sense.
Increasing a program's 'positional understanding'
has marginally improved the quality of play, and probably will continue to do so... If universal equations 'solving' chess
exist at all, it is doubtful that they would be comprehensible either to humans or computers."
We need to find a way to bring the future into the present, so the machine can decide how to focus
the search efforts along the lines of the most promising positions. We need to find a way to obtain insight into a given position
so we can accurately determine which positions are the most promising.
p.111"Strategic planning is the Achilles' heel of computer chess."
