chess knowledge vs. search depth

[CRoberson]

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

That sort of doesn't work. A chess program is fundamentally = search + material + positional eval.
The dumber version = search + material.

Since the "dumb" version is a proper subset of the "smart" version,
anything I can do to speed up the "dumber" version, I can
also do to the "smarter" version. So, why wouldn't I? This would
close the gap.

How much deeper does the "dumb" version have to go to catch the
"smart" version. That is an unknown. However, there is some data to
help make a guess. Generally, speaking (at lower Elo's) each ply is
worth 200 Elo (all else the same). So, the "dumber" version score
an Elo about 700 points below the "smarter" version. So, only another
4 ply would be needed based on that math, but I believe it will take
much more than that.

You'd have to see the games. The "pure" tactician was slowly pushed
into a strategic corner and crushed. Reminded me of the classic games
of Nimzovitch.

[smirobth]

At infinite depth you do not need knowledge

I have heard this claim many times by many people (including others in this thread), and intuitively it makes sense. But I believe it is false.

If you could search the tree completely you would never lose, but that does not mean that knowledge would be of no use; and I suspect very weak players could frequently draw against a program with no knowledge. This is because there would be nothing to stop the program with no knowledge, only infinite search depth, from playing extremely stupid moves that merely didn't lose. For example a zero knowledge program might start a game with 1.f3 and 2.Kf2 merely because it sees White can still hang on to a draw (if this is true). Then, once the program's position was so bad that it was nearly (but not quite yet completely) lost, the program would suddenly start playing like a 3000+ rated player for a few moves. But if its position started to become nearly equal it would again start playing dumb moves. So the only way the program would win would be when the opponent blundered.

[Uri Blass]

Of course you are right that knowledge can help to win against weak players even at infinite depth and I thought about the target of not losing in this discussion and not about the target of winning against weaker players.

Edit:The question in the post that I replied to it was:
"would a perfect player need any knwledge at all, other than how the pieces move?"

A perfect player by my definition is a player who does no mistakes that change the theoretical result of the game.

I do not consider not doing practical mistakes in the definition of a perfect
player.

Uri

[jswaff]

Yes, Uri already stated that so there was no need to restate.

You seemed to be offering your experiment as proof that the need for positional evaluation does not decrease with search depth. I was simply pointing out that's not necessarily true (or at least that your experiment doesn't demonstrate it's true).

I fully agree that a good balance is required to maximize the "search/eval" curve.

Anyway, we're on the same page.

Side note: seems you have improved Telepath lately, judging by Prophet's latest encounters on ICC. Congrats on the forward progress!
(But don't get too comfortable.)


--
James

[bob]

At infinite depth you do not need knowledge

I have heard this claim many times by many people (including others in this thread), and intuitively it makes sense. But I believe it is false.

If you could search the tree completely you would never lose, but that does not mean that knowledge would be of no use; and I suspect very weak players could frequently draw against a program with no knowledge. This is because there would be nothing to stop the program with no knowledge, only infinite search depth, from playing extremely stupid moves that merely didn't lose. For example a zero knowledge program might start a game with 1.f3 and 2.Kf2 merely because it sees White can still hang on to a draw (if this is true). Then, once the program's position was so bad that it was nearly (but not quite yet completely) lost, the program would suddenly start playing like a 3000+ rated player for a few moves. But if its position started to become nearly equal it would again start playing dumb moves. So the only way the program would win would be when the opponent blundered.

I think the bottom line is this: If, along _any_ line, you can't search to end-of-game, then some eval term will be required to help guide you toward positions that are not dead lost... If you can badly out-search your opponent, you can win material right and left. Unless he is fortunate enough to avoid tactical crushes. Then your additional depth won't be of benefit, and if he has superior positional evaluation (positional evaluation is just a substitute for deep searches) then he might actually be able to out-search you. Backward pawns are generally weak because they require attention. But not always. Given the choice of evaluating them or not, however, and I would always include them because more often than not, their loss is far too deep for any existing search to find.

[Bill Rogers]

Edit:The question in the post that I replied to it was:
"would a perfect player need any knwledge at all, other than how the pieces move?"
Uri

Hi I have to little chess programs that do just that. For every move they add a small random number and then select the move with the largest number, of course captures win over non capture moves. Neither one of those program can play worth a damn.
It would be extremely easy to implement on a regular program as the eval funtion is so small. Why not give it a try.
Bill

[Uri]

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.

If chess knowledge risks missing a good or even the best move, then how do Top GMs like Kasparov, Anand and Kramnik almost always find the best move without calculating all the variations which will take an awful lot of time both for a human and a program. I guess it depends on human knowledge, experience and intuition.

[Dann Corbit]

Of course you are right that knowledge can help to win against weak players even at infinite depth and I thought about the target of not losing in this discussion and not about the target of winning against weaker players.

Edit:The question in the post that I replied to it was:
"would a perfect player need any knwledge at all, other than how the pieces move?"

A perfect player by my definition is a player who does no mistakes that change the theoretical result of the game.

I do not consider not doing practical mistakes in the definition of a perfect
player.

Uri

A chess engine can play perfectly with no evaluation at all except win/loss/draw if it can search far enough.

A chess engine can play perfectly with no search at all if it can evaluate perfectly.

But the easiest way to make a chess program that plays close to perfect chess is with an excellent evaluation and an excellent search.

A perfect search is very hard to write.
A perfect eval is very hard to write.
However, a very good search coupled with a very good eval is within the realm of what a good programmer can do.
[/quote]

[Uri Blass]

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

That sort of doesn't work. A chess program is fundamentally = search + material + positional eval.
The dumber version = search + material.

Since the "dumb" version is a proper subset of the "smart" version,
anything I can do to speed up the "dumber" version, I can
also do to the "smarter" version. So, why wouldn't I? This would
close the gap.

How much deeper does the "dumb" version have to go to catch the
"smart" version. That is an unknown. However, there is some data to
help make a guess. Generally, speaking (at lower Elo's) each ply is
worth 200 Elo (all else the same). So, the "dumber" version score
an Elo about 700 points below the "smarter" version. So, only another
4 ply would be needed based on that math, but I believe it will take
much more than that.

You'd have to see the games. The "pure" tactician was slowly pushed
into a strategic corner and crushed. Reminded me of the classic games
of Nimzovitch.

If you are interested to find out you can try games at fixed depth.
My guess is that the difference in plies is not constant and 1 ply smart may lose against 5 ply dumb when 5 ply smart may win against 10 ply dumb.

Note also that only material evaluation is an extreme case of not having positional knowledge and the question is not if programs need positional knowledge but if more knowledge is better after you have basic positional knowledge.

I believe that more positional knowledge help but it is not as obvious as the example of totally not having positional knowledge relative to having positional knowledge.

[Uri Blass]

A chess engine can play perfectly with no evaluation at all except win/loss/draw if it can search far enough.

As Robin smith explained this is not correct if perfectly means beating weak players.

The engine with no evaluation except win/loss/draw can allow weak players to get easy draw(by playing poor moves that still draw and later force repeatition).

Uri