Author: Albert Silver
Date: 09:33:34 02/16/00
Go up one level in this thread
On February 15, 2000 at 01:57:10, Vincent Vega wrote:
>On February 14, 2000 at 21:12:27, Albert Silver wrote:
>
>>A little miscommunication problem here. I wasn't trying to compare different
>>programs, one with a large eval function, and another with none, as of course it
>>would be an exercise in futility for the very reasons you stated above. Instead,
>>have the same program (with a reasonably sized eval) twice, one version with the
>>eval all de-activated except the material values, and the other working as
>>usual. Now play them at a fixed depth.
>
>The same problem still exists. Knowledge isn't free, it reduces program's speed
>and therefore its ability to look deep. That's not just a theoretical issue,
>it's a real problem.
>
>>No, time is actually what is futile IMO. Time can only yield one thing: greater
>>depth in plies and everything that that ply can bring to the program. What else
>>can time bring if not depth?
>
>If you try to use fixed ply depth to compare versions of a program, you won't
>accomplish anything. I can prove that it is futile. If you take a fast
>searcher and add some good knowledge to it, it will be stronger when its
>strength is compared using plies. But its real strength may actually be weaker
>because it can't go as deep as the previous fast version.
I don't understand why you say this. Isn't that what we are debating: what one
can expect from added plies as opposed to extra knowledge, and which is more
desirable? One can't discuss different shades of grey without first establishing
what Black and White are, and that is what I'm trying to do. The first two
questions that need answering are, what do extra plies bring when they have
knowledge and when they don't. We can better understand what the knowledge
brings to a ply if we can establish what a ply with no knowledge brings.
>
>>Let's try this again: take a program and reduce it
>>so that it knows the laws of chess plus the material values only. What can it
>>learn with each and every ply? It can either calculate a forced material win or
>>a mate, but if it finds neither, it's eval will always remain zero. Always. It
>>will never learn anything about the position. It will never display a half pawn
>>advantage, nor an edge for anyside unless it sees a forced material win. In
>>fact, unless it finds either a material win or a mate, it's results will be
>>simply list of the moves that don't lead to either for it and a flipping of the
>>coin to see which gets played. It can do this at 10 plies, 20 plies, or 30
>>plies, and unless something is forced, it will know NOTHING. It's eval will be a
>>big fat +0.00.
>
>First, even that dumb program maybe be an unbeatable genius when it can look
>arbitrarily deep, because we don't know if chess isn't, as you say, forced, from
>the starting position.
Of course, if we get a computer to calculate the end-result of all chess moves,
knowledge is a completely moot point, but this argument is also moot at this
point. We are nowhere near this depth so why argue about it? Our hypothetical
friend Mindblank works within the normal limits of today's programs, and any any
arguments I present using it assume this.
>
>Second, going deeper can, and often will, improve the move selection of the
>material-only program. This simple fast program can become unbeatable. All it
>has to do is look deep enough. Of course, chess is complex enough for this
>strategy to be hopeless in practice, but this doesn't change the fact that going
>deeper does improve all search programs, even material-only ones.
>
>Note that fast-and-dumb program's evaluation doesn’t have to become good with a
>certain ply. It only has to become less bad because we are comparing strength
>improvement with deeper search, not absolute strength. Similarity good program
>has to become very good for a similar improvement. The question is whether
>getting that extra ply, for example 10'th for the slower-and-smarter version is
>better or worse than getting another ply, let's say 14th, for the
>faster-and-dumber version. Which ply is evaluated depends on their speed
>difference.
Yes, no. We now enter a new area of the argument as we are no longer striving to
find the best move, but rather to restrict the clearly bad ones. It's a fine
line but it isn't quite the same thing. Knowing that I have only 5 available
moves that don't lead to an immediate loss, isn't quite the same thing as
finding the one that leads to a win. You are absolutely correct in saying that
extra plies should help reveal what moves are not playable, but what should we
do about the ones that are? In other words we are not trying to play winning
moves, but avoiding to play losing ones. One must be careful not to confuse the
move selection with the evaluation. Suppose I find I have 4 playable moves, in
as much as no tactical refutation (forced material loss) can be found, but
neither can a forced win be found. Even if I search deeper and find that one of
the moves does lead to a forced material loss, my evaluation has not improved,
only my move selection. This will indeed improve my overall strength, but only
provided I have the tools necessary to choose the correct move of the ones left
over. You will undoubtedly note that going from a 4 move selection to a 3 move
selection represents a 25% improvement; not to be spat upon! True, but what if
the knowledge had rejected that move earlier on, for non-tactical reasons? This
now leads to a different question altogether. How much of chess is forced? I
have often heard that at many given moments, we have a plenitude of choices,
which has always been one fo the charms of chess. It allows us to express
different stylistic perspectives of the game, and this is great. On the other
hand, Boris Gelfand once said that in his opinion, almost all moves were more or
less forced, and that a player only really had an option (presumably once
leaving the berth of theory) some 2-3 times in the game, and the decision the
player took at these fulcrums represented stylistic differences. Perhaps I am
too weak to properly judge this, and I do agree that _most_ times our options
are extremely limited and many times there is really only one move, but I think
he was being a little dogmatic about the game. How does one lose then, if one
makes no mistakes? In an absolute sense, it is obvious a mistake was made (I do
not prescribe to the belief that White has a theoretical win if we ever solve
chess), but how deep? Alekhine once criticized Capablanca for not always trying
to find the best move, but contenting himself with just good moves. What then is
a best move? Does it automatically win? Probably not, but it may be the move
that provides the side with the best chances to win, in which case Lasker and
Petrosian would have something to say as they played their opponents as well as
the board. :-) A ply can bring only three things as far as I can see:
- A possible forced win, but this is only if the balance has been disrupted.
- A further restriction of the move selection by showing the moves that lose.
- A refinement of the evaluation that is ENTIRELY dependant on the program's
eval function.
In the future, I expect that we will eventually reach a stage similar to Deep
Blue, in which knowledge will be absolutely costless and will not affect the
depth of the search, which is why I find it interesting to see what one can
expect from a ply and from knowledge. The plies will be a natural consequence of
improved hardware so that work on tweaking the speed of the search will
gradually die out and further work will be done into refining the quality and
balance of the knowledge, for without quality and the proper balance of
knowledge, the knowledge can become just as much a hindrance as a boon.
>
>>>Time is an obvious constant because there is no way for a program to decide to
>>>use more time and still play another program (I guess it could use less, but for
>>>all the chess programs I know of, that would be counterproductive).
>>>
>>>>How much higher? In my opinion this probability is close to nil unless the
>>>>position has already been compromised; yet how will Mindblack achieve this great
>>>>position with random move choices?
>>>
>>>You are making an error here. The playing strength of Mindblank and CyberGM
>>>isn't in question here. The question is how much they gain by multiplying the
>>>time they have to evaluate the position. So if after 10 mins, Mindblank
>>>evaluation was weak, the question is how less-weak will it be after 20 mins, not
>>
>>See above. It will be the same.
>
>What will be the same? The probability of making the move to improve the
>position? This is obviously false, as we know that a program that looks
>extremely deep will be very strong, even if its evaluation is material-only.
This is news to me. How do we know this? I never heard of any such experiment
done.
>So
>this program's strength does obviously go up with increased plies.
>
>>You are right, you said that about CS-Tal in the beginning, and then you also
>>said that Mindblank, which only knows how to count pawns and pieces, will
>>somehow improve its evaluation with time (or greater depth as that is what time
>>brings). These two statements are in contradiction hence the lack of clarity,
>>though I can see this was involuntary.
>
>But it will indeed improve its evaluation with time!
Why will it improve it's evaluation with time? Give me a balanced position (with
a full board of pieces), and show me how it will improve it's evaluation with
greater depth if it doesn't find a forced material win or mate. If at 10 plies
it finds nothing, and at 20 plies it still doesn't see a forced material win,
how did its evaluation improve?
> And I really can't see how
>stating that is a contradiction to my liking of knowledge programs. There is no
>contradiction here.
>
>>>Even when I was just starting playing chess it was absolutely clear to me that
>>>better players not only saw deeper but also had superior knowledge of the game.
>>>Was it different for you? Did you only see tactics when you were starting out
>>>and you didn't notice any greater understanding of the game in superior players?
>>> If not why would you assume that a potentially weaker player would make the
>>>same mistake?
>>
>>Yes, and no. Chess is kind of funny in that each horizon that one reaches, shows
>>us a little more of what we don't know, and therefore a new horizon to go after
>>is revealed. Most players don't get that far along the scale and therefore never
>>realize that this process just keeps on going. Most players I've encountered
>>seem to believe that the difference between a 2200 player and a GM is only that
>>the GM is doing it better and deeper. Your comment on Mindblank learning with
>>increased depth is what led me to believe that you believed the same.
>>
>
>A statement of a trivially true
What was trivially true?
> fact led you to believe that I don't realize the
>importance of knowledge? It doesn't seem like a support for anything. Such
>reasoning seems very illogical and presumptuous.
You're jumping to conclusions. You made a comment I find dubious, and this led
me to believe something about your vision of the game. If there was no
connection between my belief and what led me to believe that (as it is there
was), then it is illogical, if not, it can just be wrong, based on an
insufficiency of knowledge. I didn't state this as a truth and realized the
possibility of error, which is why I took the time to mention it so that you
might clarify.
Albert Silver
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.