chess knowledge vs. search depth

[Peter Hegger]

Would it be fair to assume that the need for positional knowledge (or atleast positional considerations) in chess programs decreases as search depths increase? If so, does it decrease exponentially or at a constant rate?
I think Tarrasch, Nimzovich and Fred Reinfeld would be rolling in their graves if they saw the technique of a 2950+ elo program that would beat any of them effortlessly.
Todays programs seem to thrive on doubled, tripled, isolated and otherwise horrible looking pawn structures. Poisoned pawns are a special delicacy. Is it the massive search depths that are causing this utter disregard for positional considerations? Will it become even more pronounced? By extension, would a perfect player need any knwledge at all, other than how the pieces move?
Opinions?
Peter

[Uri Blass]

Would it be fair to assume that the need for positional knowledge (or atleast positional considerations) in chess programs decreases as search depths increase? If so, does it decrease exponentially or at a constant rate?
I think Tarrasch, Nimzovich and Fred Reinfeld would be rolling in their graves if they saw the technique of a 2950+ elo program that would beat any of them effortlessly.
Todays programs seem to thrive on doubled, tripled, isolated and otherwise horrible looking pawn structures. Poisoned pawns are a special delicacy. Is it the massive search depths that are causing this utter disregard for positional considerations? Will it become even more pronounced? By extension, would a perfect player need any knwledge at all, other than how the pieces move?
Opinions?
Peter

Your assumption is not correct.
At infinite depth you do not need knowledge but we have no infinite depth and at finite depth knowledge help and the same for search depth.


I do not know what do you mean when you say that
"Todays programs seem to thrive on doubled, tripled, isolated and otherwise horrible looking pawn structures."

You should give examples because maybe the programs know to evaluate that other positional factors are more important than these pawn structures and it clearly means that positional considerations are important.

Uri

[CRoberson]

I've had this debate before and no mater how much I explain the
theory people just don't get it. So, here are the results of an experiment
I ran a few weeks ago.

I ran my engine (Telepath) against a group of 5 engines at 8 games
per engine. It scored 50% (20 out of 40). I then ran another
identical version but the positional knowledge was removed. The new
dumbed down version ran 10x faster and reached an extra 3 ply deep
in the same amount of time on my benchmarks. I reran the same
matches with the same group of opponents. In highly tactical positions,
the new dumbed down version reached depths of 5 and 6 extra plies.
So, by your argument the dumb faster version should perform much
better.

Instead of gaining 20 out of 40 or better, it score only 1 draw and
39 losses. Clearly, the positional knowledge was useful.

[jswaff]

I've had this debate before and no mater how much I explain the
theory people just don't get it. So, here are the results of an experiment
I ran a few weeks ago.

I ran my engine (Telepath) against a group of 5 engines at 8 games
per engine. It scored 50% (20 out of 40). I then ran another
identical version but the positional knowledge was removed. The new
dumbed down version ran 10x faster and reached an extra 3 ply deep
in the same amount of time on my benchmarks. I reran the same
matches with the same group of opponents. In highly tactical positions,
the new dumbed down version reached depths of 5 and 6 extra plies.
So, by your argument the dumb faster version should perform much
better.

Instead of gaining 20 out of 40 or better, it score only 1 draw and
39 losses. Clearly, the positional knowledge was useful.

The only thing that proves is that the extra plies you got were not sufficient to overcome what you lost in positional knowledge. It does not prove that the need for positional knowledge does not decrease with search depth. Clearly it does - if you could search the tree completely you wouldn't need *any* positional knowledge at all (do you need heuristics in tic-tac-toe?).

[Uri]

Chess knowledge (like quiescent search) speeds up the program and makes it more reliable but i'm no expert. A Top GM like Garry Kasparov or Vishy Anand searches only 3 moves per second but still finds the best move thanks to knowledge and intuition.

[CRoberson]

The only thing that proves is that the extra plies you got were not sufficient to overcome what you lost in positional knowledge. It does not prove that the need for positional knowledge does not decrease with search depth. Clearly it does - if you could search the tree completely you wouldn't need *any* positional knowledge at all (do you need heuristics in tic-tac-toe?).

Yes, Uri already stated that so there was no need to restate. As I said,
I've debated the theory before and some just don't understand the
concept with two extremes - "with perfect eval there is little need
(if any) for search and with infinite search depth there is no need for
eval other than draw and checkmate". Simple two ended theory, but
I've seen people struggle with it. Thus, I delivered the results of a
simple experiment.

The point I tried to make was this - increased depth isn't always better.
However, assuming all else remains the same it typically is but not
always there either (ie. too much pruning ...).

That is part of the fun of computer game competitons. Since we don't
have full search in games like chess, shogi and most definitely go,
we are forced to seek a balance between search and eval. The neat
part is when two programs compete whose authors have chosen
different balances in that equation.

[CRoberson]

Chess knowledge (like quiescent search) speeds up the program and makes it more reliable but i'm no expert. A Top GM like Garry Kasparov or Vishy Anand searches only 3 moves per second but still finds the best move thanks to knowledge and intuition.

Actually, chess knowledge often slows down the search. That is
chess knowledge added to the position evaluator. Chess knowledge
used in the search for pruning can speed it up.

At the end of every "search line", one has to evaluate the resulting
position. This is where many people make the mistake of chosing a line
because of a nice combo but falling into a bad position. Computers and
GM's don't make the mistake of not evaluating the resulting position
before making the first move in the sequence. That is one use of
"chess knowledge".

The other is when the person knows that in a given position there are
only 3 good moves. Then he has to search lines that start with each
of those moves. At that point, he has pruned the search by only
considering 3 starting moves out of typically 40. This risks
missing a good or even the best move.

[Peter Hegger]

the new dumbed down version reached depths of 5 and 6 extra plies.
So, by your argument the dumb faster version should perform much
better.

Instead of gaining 20 out of 40 or better, it score only 1 draw and
39 losses. Clearly, the positional knowledge was useful.

Hello,
I agree that 5 or 6 plies would not be sufficeient to offset positional knowledge. However, for arguments sake, lets say the hypothetical difference was on the order of 20 plies. If the dumbed down version was able to score 1 draw at 5 or 6 plies difference, would it not score better with the increased depth advantage. And wouldn't the difference become more pronounced as the the depth advantage was increased?
Peter

[bob]

Would it be fair to assume that the need for positional knowledge (or atleast positional considerations) in chess programs decreases as search depths increase? If so, does it decrease exponentially or at a constant rate?
I think Tarrasch, Nimzovich and Fred Reinfeld would be rolling in their graves if they saw the technique of a 2950+ elo program that would beat any of them effortlessly.
Todays programs seem to thrive on doubled, tripled, isolated and otherwise horrible looking pawn structures. Poisoned pawns are a special delicacy. Is it the massive search depths that are causing this utter disregard for positional considerations? Will it become even more pronounced? By extension, would a perfect player need any knwledge at all, other than how the pieces move?
Opinions?
Peter

I am not sure it is obvious at all, except at the limit where we search to find a mate, so that no evaluation of any kind is required. When Cray Blitz was searching depth=8 in 1981, we were happy with the depth and were working on the evaluation. Today, hitting depths of 18-20, I am _still_ working on the evaluation. So while there might be a point where as you go deeper, you eliminate some evaluation requirements, it comes about slowly. For example, I have not evaluated pins in years, nor forks, because I let the search handle those. But pawn structure is so far beyond what a search can resolve, I am adding to that weekly...

[bob]

Would it be fair to assume that the need for positional knowledge (or atleast positional considerations) in chess programs decreases as search depths increase? If so, does it decrease exponentially or at a constant rate?
I think Tarrasch, Nimzovich and Fred Reinfeld would be rolling in their graves if they saw the technique of a 2950+ elo program that would beat any of them effortlessly.
Todays programs seem to thrive on doubled, tripled, isolated and otherwise horrible looking pawn structures. Poisoned pawns are a special delicacy. Is it the massive search depths that are causing this utter disregard for positional considerations? Will it become even more pronounced? By extension, would a perfect player need any knwledge at all, other than how the pieces move?
Opinions?
Peter

Your assumption is not correct.
At infinite depth you do not need knowledge but we have no infinite depth and at finite depth knowledge help and the same for search depth.


I do not know what do you mean when you say that
"Todays programs seem to thrive on doubled, tripled, isolated and otherwise horrible looking pawn structures."

You should give examples because maybe the programs know to evaluate that other positional factors are more important than these pawn structures and it clearly means that positional considerations are important.

Uri

There are lots of cases where top GM players criticize a pawn push by a computer, but when they try to demonstrate how bad it actually is, they fail. It reminds me of the old KR vs KQ strategies believed to be true for a hundred years or two, namely that keeping the rook close to the king is the best defense. We now know that is nowhere near truth, as Ken Thompson's "the great flying rook" explanation showed once he did the KQKR endgame table...

As the games get deeper, traditional pawn structure ideas slowly change as the computers show that the ideas were not the truth, the whole truth, and nothing but the truth...