Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: DB will never play with REBEL, they simple are afraid no to do well

Author: Ratko V Tomic

Date: 22:24:48 10/14/99

Go up one level in this thread



>> These small effects will thus not add up coherently
>>into a winning strategy as they do in high level human play.
>
>Depends on the program. My current program design calls for long term plans
>where "undoing" previous plans must be justified by a tactical advantage.
>Basically, it will have multiple PVs and select it's PV based on how well it
>follows the current plan. Hence, your concept of random optimal moves should be
>minimized.
>

Well, that's interesting. What's your program's name and is it a playing program
as yet? (I have a similar project, but it is still in its early infancy, much
more research oriented than for a practical play.)

I was talking in this thread about current state of the art alpha-beta
searchers. The deeper they look the greater is the percentage of junk positions
they weed through, and that is an exponential convergence toward 0 percent of
useful positions examined. Yes of course, they see more useful positions too,
just because they see so many more positions in absolute numbers, but the net
efficiency in useful processing still drops exponentially toward zero with the
depth. And this type of phenomenon, in any kind of evolving or competing system,
can only mean one thing ahead: a dead end, an extinction.

As far as the ultimate outcome, yes programs will be completely superior to
humans in chess, but not the present dominant kind of programs. These present
ones will go the way of the dinosaurs. While the alpha-beta search will remain,
it will be a lower layer in a multilayered control, where the higher layers
operate with plans and goals and guide the lower layers into much greater
percentage of useful work than presently. Botvinnik's project was an early
attempt at these higher layers in the future schemes. Even though his work is
being ridiculed by many of the current chess programmers and theoreticians, the
mammals must have looked ridiculous to the powerful dinosaurs in their heyday.
First clunky car prototypes looked ridiclous to the carriage and railroad
owners, first airplanes looked silly to baloonists.

>
>Stupid is a bad choice of a term here. I meant that they play sub-optimal
>throughout the game. They may make 60% of the most optimal moves, but 40% of
>their game is based on sup-optimal moves and the reason they win is due
>to their opponent making even more sub-optimal moves. This is not meant
>as a dig at humans. Rather it reflects the reality that the game is
>not totally understood and not perfectly played.
>

In some aesthetic (or other arbitrary, subjective) sense, it may be imperfect.
But give a GM an endgame with an extra piece, and however imperfectly he will
generally (short of rare pathological positions) win a full point from anyone,
program present or future, even the table-base with perfect knowledge. He needs
only to find  "good enough" moves, not the "perfect" moves (perfect by some
criteria or value scale which is outside of the rules of the game).

If you can get at most a point for a game, you can't call genuinly suboptimal
any technique that gets the full point. Since optimal or suboptimal can only be
referred as with respect to some reward function, the reward function in chess
is flat at the checkmate. Pretty or ugly checkmate, deep one or a cheap shot,
slow or quick, one point remains one point. All other arbitrary criteria one may
talk about (when using "optimal" as you did) are outside of the rules of the
game, and have no effect for rating or tournament/championship rank. A program
may be collecting and happily counting its evaluation function pluses all game
long and fall short only once when it really counts.


Given that the objective of the game is not the shortest or prettiest, most
sacrificial (or whatever else) checkmate, but any checkmate, it seems to me
absurd to have enormous tablebases with "optimal" wins, when a much much smaller
data base with "sub-optimal" wins (e.g. in the form of extracted rules and
patterns) can gain as many torunament or rating points. Actually if two programs
have the same hardware, CPU speed and amounts of disk space, the one with
"sub-optimal" database will win more points since it will be able to keep more
different endgames (based on "good enough" choice) in the same disk space.
Additionally, dropping the superficial "optimality" criteria (the shortest win
requirement, which doesn't contribute to the number of wins), the pattern
matching algorithms may be able to take advantage of "suboptimality" and work
faster or find the match earlier in the tree (via rules, like the those used for
pawn races, kings opposition manouvers, rook checking directions, etc), and thus
gain even more real points for the "sub-optimal" program.

The same applies to the other phases of the game, i.e. a program which knows how
to use well the relaxation of the goal from the "looking for the best move" to
the "looking for the good enough move" will score more points than the
"perfectionist" (in its own mind, only) program, playing on the same hardware.



> Although this theory sounds promising on the surface, I do not think it is
> correct. The GMs are losing more and more games at serious times
> exactly because the programs discover tactics which the human GMs
> misses. Since it seems obvious that programs are getting stronger
> and stronger as time goes on (regardless of whether the reason for
> this is faster systems, better heuristics, or a combination of the two),
> it also seems obvious that the humans are not getting much better
> at anti-computer strategies.

How many books and papers and researchers deal with computer chess and how many
with human anti-computer strategies? Well, as you can see, computers have had a
virtually completely free ride, with no scientific (or any other kind) research
being directed in the human strategies for anti-computer play (again I am
talking here about the present type of essentially brute-force searchers).

That doesn't mean these strategies don't exist. The positions certanly exist, in
good, usable, percentage, where a deeper search gives worse move (as the example
of a good percentage, 20-30 percent perhaps, of games won by the N-ply searcher
against the identical N+1 ply searcher shows). So the comprehensible strategies
and opening lines likely exist which tend to steer the game toward these kinds
of positions, and humanly comprehensible patterns likely exist which can pick
them out and trigger appropriate strategies. Of course, once these methods come
into focus and develop, the balance will tip back, until the next generation of
the truly smarter programs replaces the present alpha-beta searchers.



>The problem is one of the algorithms being so sophisticated and the systems
>being so fast that mere bean counting will be capable of playing a game where a
>human just cannot keep up the grueling pace for an entire game.


Well, a mere "bean counter" searching N+1 plies still loses at least 20% of time
to N-ply "bean counter", even though it outcalculates it, and its own mind (i.e.
by its own evaluation points), outsmarts it in every single move. The thing is
that what N+1 "bean counter" is counting isn't worth (often enough) what it
thinks it to be worth. If I outsmart you in every single move, I should win
every single game. So the N+1 ply program is deluded in its belief that it is
outsmarting every move the N-ply program. It is counting things that
_ultimately_ don't count.


>It is not a question of whether this will happen. It is a question of when.

While I agree with the conclusion, there is an important question: how, by what
type of programs will this be achieved. Not by the brute force searchers of
present-day variety (whatever minor refinements they may gain in hardware speed,
move ordering, bean counting terms, etc).


> And the GMs who have analyzed the playing style of programs have yet to
> come up with this panacea of anti-computer playing style that you are
> talking about. If they haven't come up with it yet (when chess is the
> life blood for some of those guys), it seems extremely improbable that
> they ever will.

While chess may be the lifeblood, it's not the computer chess that is as
important to the top players (other than in brief jolts of manic activity driven
by the PR motives of computer companies). They have analyzed it from the
perspective of playing another human-like oponent, who is just a very strong
tactician, but still as if there is another mind of similar kind to himself
behind the moves. That is a very superficial and misleading model of the current
programs. Their key weaknesses are much more deep rooted and much more fatal for
their play. When the real interest emerges (unrelated to occasional exhibition
match and PR gimmicks) in the IM/GM world in beating the present type of
programs, that field will develop and you'll have as many books and scientific
papers on that topic as you do on the currently prevaling dogmas in chess
programming. And the free ride for the bean counters will end there.

Another possible path may be that the current "bean counters" will go extinct,
displaced by the more advanced kind of programs, well before the interest in the
chess community in beating the "bean counters" ever picks up the momentum (if it
ever does). In that case your scenario of computers eventually taking over
(without any major comeback of the humans) will outwardly go as you described it
(but not exactly for the reasons you picked, i.e. for the ever stronger
tactics).



This page took 0.01 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.