Even crazier

[syzygy]

This reminded me of a post by Chris Whittington, a most remarkable one.
See here: http://www.talkchess.com/forum/viewtopi ... 77&t=44969

He's effectively arguing that top engines can't compete with full width searching humans. Now show me such a human.

[carldaman]

This reminded me of a post by Chris Whittington, a most remarkable one.
See here: http://www.talkchess.com/forum/viewtopi ... 77&t=44969

Wow, I remember Chris W. from the old days, and his constant berating of so-called "bean-counters", but I hadn't seen this post. However, the specific idea of 'incestuous', in-bred engine-only testing that produces illusory rating gains, which don't translate into better play vs human players, has been on my mind a lot recently. I wish C.W. could return to further contribute to this thread. It would be very interesting.

[carldaman]

This reminded me of a post by Chris Whittington, a most remarkable one.
See here: http://www.talkchess.com/forum/viewtopi ... 77&t=44969

He's effectively arguing that top engines can't compete with full width searching humans. Now show me such a human.

I don't think the current argument is that humans play stronger chess than the top engines, but rather that top engines can underperform badly against humans of moderate strength (2000-2300 FIDE). Lyudmil Tsvetkov, a player of such strength, has produced quite a few recent examples to support such a thesis.

[carldaman]

Hi Carl.
Thanks for the precious support!

2 against the pack. Now I am not afraid any more.

My only further comment to this would be that you can made an engine both appealing and strong, both human-like and engine-like, just like humans might learn some of the tactical tricks of engines.

Well, I think it's wonderful that you accomplished the hard part of getting tangible positive results against Stockfish, exposing real weaknesses in its play that top engines cannot normally exploit (due to having similar weak points themselves).

It is because of this that the onus is on the programmers to make the engine better, and our theories regarding engine play have even more credibility, since they can be backed up by actual results.

I keep thinking that if were an engine developer, I would start out with an entirely different paradigm. My focus would be on creating a planning-based human-friendly program, where actual maneuvers and plans would be carried out, even if they could be refuted by strong human play. I say 'human play' because I would only test and evaluate such an engine against human opponents, which is not so hard to do, as an engine can play on a server such as ICC, and only be allowed to play humans.

I would pack as much knowledge into the engine as possible without fear of 'slowing' it down. Any improvements would be assessed based on its play and results against human competition. Tactical strength would be of secondary importance, but not altogether neglected. The burden of proof would be on the humans to refute the engine's plans. Human feedback would be more than welcome -- and actually essential to further improvement.

The vast majority of developers shudder at the idea of weakening their engine tactically, being slowed down by 'unnecessary' knowledge. Such ideas are anathema to so many (my apologies to those who think differently, of course). Most engines have an over-abundance of tactical strength that could partially be traded off for better play where the engines are weak, but God forbid an engine should lose some precious Elo points! If developers begin to overcome and manage this great fear, then we may indeed see the dawn of a new era in computer chess, where something that resembles true AI can come to the fore, that can benefit chess players more than what we have right now.

The kind of program I envision could be a great sparring partner for humans, simulating other humans better than it can be done at the moment. Play against such an opponent would be a truly fascinating experience. As it becomes stronger, it could also become a great analysis tool. To get there, we must start in the right place, and let go of the fears of our product being ridiculed as 'weak' -- such a relative term, after all.

Regards,
CL

[syzygy]

I don't think the current argument is that humans play stronger chess than the top engines, but rather that top engines can underperform badly against humans of moderate strength (2000-2300 FIDE). Lyudmil Tsvetkov, a player of such strength, has produced quite a few recent examples to support such a thesis.

It's not difficult to produce a winning game against any engine, but under equal conditions it's not so likely to happen.

[syzygy]

It is because of this that the onus is on the programmers to make the engine better, and our theories regarding engine play have even more credibility, since they can be backed up by actual results.

What are the actual results? That positions have been found that are misevaluated by most engines?

Engines will get better, but there will always remain positions that will be completely misevaluated. And there is no shred of evidence that adding a rule for each position found is going to make the engine play better against either itself, other engines or humans.

I keep thinking that if were an engine developer, I would start out with an entirely different paradigm.

If you had been an engine developer, you would probably not say this, because you would not have a clue how to go about it.

You can not just decide to dream up an "entirely different paradigm". The world does not work that way.

But naive non-developers will keep repeating your dream even 500 years from now when engines are playing at 5000 Elo or so.

[carldaman]

I don't think the current argument is that humans play stronger chess than the top engines, but rather that top engines can underperform badly against humans of moderate strength (2000-2300 FIDE). Lyudmil Tsvetkov, a player of such strength, has produced quite a few recent examples to support such a thesis.

It's not difficult to produce a winning game against any engine, but under equal conditions it's not so likely to happen.

Even with the extra time that Lyudmil allotted himself, a true 3000+ entity should have performed and played better. Lyudmil also showed that he can overperform vs Stockfish far more than once (that is, with a certain regularity).

[carldaman]

It is because of this that the onus is on the programmers to make the engine better, and our theories regarding engine play have even more credibility, since they can be backed up by actual results.

What are the actual results? That positions have been found that are misevaluated by most engines?

Engines will get better, but there will always remain positions that will be completely misevaluated. And there is no shred of evidence that adding a rule for each position found is going to make the engine play better against either itself, other engines or humans.

I keep thinking that if were an engine developer, I would start out with an entirely different paradigm.

If you had been an engine developer, you would probably not say this, because you would not have a clue how to go about it.

You can not just decide to dream up an "entirely different paradigm". The world does not work that way.

But naive non-developers will keep repeating your dream even 500 years from now when engines are playing at 5000 Elo or so.

OK, Ronald, I take it you will stick with what you know best. Nothing wrong with that.

I'm curious, still, to know if anyone much more skilled than myself has attempted the approach I'm envisioning. Is it accurate to say that it cannot be done, (or done well) ?

My problem is with the way so many developers are so opposed to adding chess knowledge because it slows down their program. Elo is not everything to everyone.

[Lyudmil Tsvetkov]

Two things.

1. Why try to invent a NEW pawn vocabulary? There are several good books on pawn structure. Kmoch comes to mind.

2. I don't like terms like "semi-backward." Why not just do as most upper-level players do and consider weak pawns. A backward pawn is weak. An isolated pawn is weak. An artificially isolated pawn is weak. Kmoch defines all of those. Ditto for majority, crippled majority, mobile majority, immobile majority, etc...

But these new terms are not going to catch on very well since there is already a lot of pawn structure analysis in GM books...

Mr. Hyatt, I respect you very much for your enormous contribution to computer chess, but, what would you suggest instead? Stick to the notions people invented 30-50-100 years ago and not progress any more, simply because we regard those doctrines as a dogma?

Do you really suggest that a scientific mind doing research some 100 years ago would really surpass persons that would inquire into interesting aspects of knowledge that probably exists, but is still unveiled? It would be really funny if it is like that. Actually, most branches of science and even art have already attained levels where there are very few substantial things to do, but, fortunately, computer chess is not one of those.

Do you realize the enormous progress being made in computer chess every year, including in terms of elo increase to the point where many programs already surpass the best humans? How much of this progress do you think is due to implementing eval terms in general, and, if you want to make a distinction, in applying existing and tinkering with new terms, previously unused? I guess eval would be responsible for 1/3 of all progress, including search and hardware. That is not negligeable.

I myself do not have any ambitions at all, I would be glad to help a bit some engines, if possible, as engines, including Crafty (thanks for this software!), have helped me quite a lot in my development. But is not it evident that some eval terms are really outdated? You can not do a 21-st century engine with concepts from the middle of the 20th century. Could you tell me what auothors of renown have suggested pawn concepts like backward-fated, unopposed, apex pawns, proximity of peak pawns to the enemy king, etc.? Do you think they are not valid concepts? Do they have to be suggested by Hans Kmoch in a paper edition to be valid?

I would suggest to do the following: try measuring the impact of, say, apex pawns, in Crafty, very easy to do, and see what happens. I suppose, if correctly implemented, with no redundancies, you will see some elo increase in Crafty, it might be 10 elo, but will be there. Why reject the existence of concepts that are useful?

Best, Lyudmil

Sorry, but nothing in your "new terms" is new. Read some good books on pawn structure, and you will see that you have not identified anything new at all...

Most authors I have read discuss your concepts. The "apex" is a part of a common attacking theme. Which is often modified by the direction of the pawns. For example, white pawns at d3,e4,f5 are in a chain pointing toward the enemy king, helping an attack, while black pawns at f6, e5, d4 also form a chain, but it points in the wrong direction.

As I said, using new terms for existing ideas doesn't help at all...

And why in this case engines in general, even all top, do very badly in terms of understanding chains?

Where are apex, uopposed and backward-fated pawns specifically mentioned, just for the record?

I don't consider your idea of "unopposed" to be relevant. You gave an example of black pawns on a7, b7 and c7, and white pawns on b2, b3 and c2. That black pawn on a7 is NOT any sort of "positional advantage" whatsoever. It does not threaten to create a passed pawn. It does not threaten to disrupt white's already disrupted pawns.

So what, exactly, is the point. For backward or "backward fated", the ideas are well known. And fall under the classification "weak pawns".

Whether chess books cover the topics has NOTHING to do with whether engines understand the concepts. There is no connection. My computer doesn't read chess books and learn how to play better.

[d]6k1/1pp5/1p6/8/PPP5/8/8/6K1 w - - 0 1
Do not you see, an unopposed pawn could be worth because of its ability to advance freely. Instead of assigning just penalty for double, you could assign smaller penalty for double and also a small bonus for unopposed, in this way the chess logic will be better reflected. You never know, in some cases the important thing might be that the enemy pawns are double, structurally deficient and vulnerable, and in other cases what will matter will be the ability of the unopposed to advance.

[d]6k1/pp2pp2/1p2p3/8/PPPPPP2/8/8/6K1 w - - 0 1
Again, neither c4, nor d4 are potential passers, but in tandem they are important. Quite probably, white will win this position, and also similar structures with more pieces on the board, and that will not be because specifically of the black double pawns (some even would dismiss their importance), but mainly because of the c4 and d4 unopposed pawns. Those pawns matter, they fulfill a certain logical function, and should be scored in a way. How otherwise are you going to catch the momentum of the position. You need a term to know in more complicated positions what the real chances of success are. The more terms you have, the better you will assess those chances.

How you call those pawns is not important, the important thing is that they are real.

Concerning weak pawns, frankly speaking, I still have not been blessed with a chance to be able to play against a program that even slightly understands backward-fated pawns. All engines I know completely misunderstand those pawns. And they are very important, with values going as much as twice those of backward pawns. If you have a single category, you need to distinguish between values. Also, a backward-fated pawn, part of the king shelter, will bring you a loss, not very much so a backward one.

Computers do not read chess books, but their authors do. And it is really surprising that with such a quantity of books read no engine author has succeeded to render successfully in his software the ideas supposedly encountered in those books, be it only weak pawns.

[Lyudmil Tsvetkov]

This reminded me of a post by Chris Whittington, a most remarkable one.
See here: http://www.talkchess.com/forum/viewtopi ... 77&t=44969

Hi Joerg, excellent post, I had missed that.

My vision of engines of tomorrow, without knowing what I am talking about, would be much heavier and refined evaluation, combined with much higher selectivity, but smart selectivity, prune completely entire branches, but also consider completely other branches.

Concerning tactics and strategy, I do not know what one and the other means apart from conventional use. It is maybe obvious that you can not separate them. Modern engines are going after perfection, the key would be to be as precise as possible in both evaluation and search. And that supposes a multitude of small steps.

And, please, do not mention to me any more cypress tree searchers