Computer Chess Club Archives


Search

Terms

Messages

Subject: Philosophical thoughts: Probability Eval vs. Conventional Evaluation

Author: Stephen A. Boak

Date: 00:25:19 01/17/02

Go up one level in this thread


Somewhere below in this thread:

On January 16, 2002 at 07:48:51, Russell Reagan wrote:
>
>Pawns, percent, queens, dollars, monkeys...all units will work equally well if
>you can compute an evaluation in terms of them from a chess position.
>

1. It is trivial to play semantic games with unit of measurement names.  One
measurement works equally well compared to the same measurement, no matter by
what name the rose (same version) is called.

2. Different measurements (evaluations) do not work equally well (in general).
I'm not simply talking about translating one measurement's scale by adding a
fixed amount, or by multiplying the measure by a fixed amount.

3. The question is whether one kind of measurement is better than another kind
of measurement for chess analysis purposes.

3A. Chess analysis purpose is to determine what is the best move, or which of
several move possibilities is better--in 'unsolved' positions.  Since chess is
not solved, the purpose is to make this determination in an uncertain way (yes,
probabilities are involved in one's judgement--flawed though it may be at times,
and accurate though it may be at times).

4. Since chess is unsolved, selecting a best move is unsolved (in general); and
even more so unsolved (in general) is selecting the move with the best
'chances'.

5. In an unsolved position, the probability of winning (or scoring expectancy)
is not based on simply solving the chess tree from the given position, and
computing and rolling up somehow the min/max percentage of lines score W, D or L
among all the branches (even if that could be done!).

Ideally, it is also based on the probability that the opponent (and every one is
different, whether human or program) will more likely lose his way among the
unsolved tree branches than the evaluator (program we ideally want to have--that
computes perfect scoring expectancies).  And this is *impossible* to program (my
hyperbole; tongue in cheek, of course--one can always try!!!).

If our program A is very good at judging the relative imbalances, say, of two
knights vs a bishop and a knight in a closed position, versus the opponent
(human or program B), then steering the game into such positions where the
knights are evaluated as better than the opponent's knight & bishop (or the
alternative) would be smart.  Not simply because the chosen configuration is in
fact better, but because the opponent likely has far less ability concerning how
to manage his side with N + B.

6. In addition, the goal of seeking a win at all cost (to get a prize, for
example), may promote taking a riskier game path with lower overall scoring
expectancy, simply because the chance of winning (via the opponent going wrong)
is a bit higher.

7. If you were Tal, would you be content with, say, a 0.25 (1/4) pawn advantage
(computed conventionally) with an uphill middlegame battle to try to convert
that small edge into a definite win, or instead with a 1 pawn disadvantage (sac
of minor piece for two pawns) where the resulting chaos catered to your style
and inclinations and sense of what the opponent might not be able to handle as
well as you?

7A. I believe that you have to take *some* risks in order to win an otherwise
dull game (in general) when at least one successful defensive path for the
opponent has a high probability of being detected and taken without mishap.

7B. A simplistic probability evaluator might look at several of 35 move choices
and determine that for one, the opponent must make 1 correct move out of 40
possible, else would lose.  For another chosen move, the opponent might have 2
out of 40 possible moves that avoid immediate lost--but those moves might be
extremely difficult to visualize and calculate, even if the thought occurred to
the opponent that such moves might be candidates.  The 'materialistic'
probability evaluator program (heh, heh!) would use 'pure' math to chose the
move that gave its opponent only 1 (obvious, but singular move) out of 40 move
chances of correctly defending; whereas the more sophisticated (Tal-like)
probability evaluator program would make the move that gives the opponent 2
(extremely difficult, but not, however, singular) of 40 reply
opportunities--because the 2 successful defensive moves are so very difficult to
see.

Conclusion--Thus we see that a probability based evaluator is also subject to
'material' shortcomings (heh heh!), which it was designed to avoid!, and
furthermore, it may well need another bootstrap evaluation function (Maxwell's
demon!) to determine if lesser probabilities are indeed better than greater
probabilities (ho ho!) for pursuing victory (or higher scoring expectancy) on
the chessboard!  :)

7C. Keeping some tension in the game, keeping some imbalance in the game (with
no perfect assurance that the imbalances will favor your side in the ultimate
analysis) is necessary to lay *some* land mines about the position that your
opponent may step into.

It often takes several moves that entail some risk (add or keep unsolved
complexity in the game), to keep the game complicated enough that your opponent
steps off the edge and loses via a mistake.  It is a fine art for a good player
to manage his game in just this manner (when desired or necessary to seek a
win), judging that his own skill in handling the position is likely better than
his opponent's skill.

8. My bottom line is that it is hard enough to program a good playing program
using conventional evaluation (basically in terms of a pawn unit) that can
evaluate (judge) positions with imbalances very well in all manner of possible
openings, middlegames & endgames, let alone one that evaluates in *probability*
terms with even greater success.

8A. What are the signposts along the way, in a string of unsolved positions (and
move choices), to guide the probability evaluator?  Perhaps the best signposts
are provided by the largely material scored, conventional evaluator, outputting
imbalance judgements in terms of net pawn units--for steering the game toward
positions of greater win or scoring expectancy likelihood?

How could you collect data on enough unsolved positions to tune the probability
evaluator?  How would you categorize the obtained subset of possible (or even
likely) positions in order to determine progress--since the vast majority of
chess games are unique.  How would you know the chance of success (win, or
maximized scoring expectancy) a priori?

These are tough philosophical questions that make envisioning a probability
evaluator that is better than the current conventional evaluator hard to do.

8B. Random programming won't do the trick.  But, carefully planned strategic
ideas in the mind of the programmer may (*may*, I say) lead to a breakthrough.

8C. Hoisting oneself by ones bootstraps via assuming you can simply evaluate
winning or scoring chances in a 'non-material' way not tied to pawn units, so
why doesn't somebody do it, is a bit naive.  It is a subject worthy of
contemplation, but making quick assumptions is fraught with conceptual
difficulties that need to be resolved.  Aye, there's the rub!

8C1. The vast majority of won games we have seen are simply and effectively
evaluated as won games--solely via material scoring (a pawn up being often the
simplest such).  Sure there are exceptions, and surprisingly many of them as
many creative GMs have shown, since chess is a huge universe of possible
positions.  But don't you have to balance the chances of certain imbalances
being worth more or less than the easily counted pawn units that speak
eloquently for their own chances?  This is no trivial matter.

Just some thoughts,
--Steve




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.