Stoofvlees II d2 is Killing Stockfish in TCEC

[chesskobra]

It is obvious that chess engines, self-driving cars, ... still have blind spots, and chess engines sometimes don't see what a low rated player may see without calculation. So I agree that an engine that knows more patters (either by discovering them or as in HCE) would play better chess than an engine that does not see these patterns. But I still do not see how an engine would play perfect or near perfect chess if we found more patterns. Eventually - very likely - someone like twoforce who dedicated his life to finding deeper patterns would come to the conclusion (if he is lucky) that almost every position has its own distinct pattern. It is somewhat like this. If you have a set of objects (e.g. squares on the board or a set of pieces of different kinds) and you choose some combinations from the set, you would not find many patterns in it (every combination is a pattern in itself) unless the set has a very special structure (or symmetries?) (e.g. if all objects are identical). You may find some combinations with a common pattern, but the number of distinct patterns would grow exponentially with the size of the set, like the number of subsets of a set, unless there is a lot of structure. It is very unlikely that such a structure exists in chess simply because of a variety of combinations of squares and a variety of pieces. Many combinatorial objects have this issue, which is why there is 'combinatorial explosion' in most situations.

[towforce]

2. To find out which chess engines are the best. For this, you really need bigger boards (12x12 might be a good starting point).

Translation: in order to find the best chess engine, we have to play some game other than chess. Doesn't make sense.

Draw death is on the way for computer chess. Once an engine becomes unbeatable, it can't really get any better.

A game very similar to chess but on a larger board (call it "Engine Chess") would handily resolve this by exploding the branching factor (the aspect of Go that made that game so difficult for engines to master). The engines that would then win would be the ones that know the game well - not the ones that process the game tree the fastest.

[towforce]

It is obvious that chess engines, self-driving cars, ... still have blind spots, and chess engines sometimes don't see what a low rated player may see without calculation. So I agree that an engine that knows more patters (either by discovering them or as in HCE) would play better chess than an engine that does not see these patterns. But I still do not see how an engine would play perfect or near perfect chess if we found more patterns. Eventually - very likely - someone like twoforce who dedicated his life to finding deeper patterns would come to the conclusion (if he is lucky) that almost every position has its own distinct pattern. It is somewhat like this. If you have a set of objects (e.g. squares on the board or a set of pieces of different kinds) and you choose some combinations from the set, you would not find many patterns in it (every combination is a pattern in itself) unless the set has a very special structure (or symmetries?) (e.g. if all objects are identical). You may find some combinations with a common pattern, but the number of distinct patterns would grow exponentially with the size of the set, like the number of subsets of a set, unless there is a lot of structure. It is very unlikely that such a structure exists in chess simply because of a variety of combinations of squares and a variety of pieces. Many combinatorial objects have this issue, which is why there is 'combinatorial explosion' in most situations.

I have no doubt that SAT (Boolean satisfiability) does not have enough patterns to be able to solve it quickly via pattern recognition. In chess, though, the evidence for good patterns existing is overwhelming:

* GMs can see many positions and understand them in a heartbeat (I have personally witnessed this)

* Chess is full of symmetries

* there is a lot of dependence between the variables (as opposed to "independent variables") in chess

* the book "Chess Skill In Man And Machine" cites a study that shows that to become a GM, you need expert knowledge of 50,000 chess patterns. This is a TINY number in comparison with the set logic ("the number of subsets of a set") you were talking about - and there is ABSOLUTELY NO REASON WHATSOEVER to believe that human GMs have managed to encode the best chess patterns possible

* nature is an absolute master of finding simple patterns in complex systems with many more variables (and variables that don't interact in simple ways like chess pieces): a fruit fly's brain has only 120,000 neurons - and yet it can learn skills and has mastery of a large range of skill sets (including flight control, which most humans cannot do). A "reasonable stab" at fruit fly control using today's tools would probably be a software project of multiple gigabytes

It's ridiculously obvious to me that chess has some big simplifying patterns. I acknowledge that I don't have actual proof of this right now - but for myself, I don't need proof. I don't need anything more than a glance at the system for complete faith in what I'm saying.

[chesskobra]

Of course, GMs know many more patterns. But we also know that GMs are nowhere close to beating even the current, less than perfect, top chess engines.

Have you found any new simplifying pattern that is not known, for example, an endgame principle that is not in any of the books by Fine and Benko, Nunn, Müller, etc.? If there are some big simplifying patterns for chess, what would stop us from having such patterns for NxN chess?

You had mentioned the game of Nim in another post. Nim is different. It is an impartial game, and impartial games have a nice structure, as characterised by the Sprague-Grundy theorem. Such simple structure is not known for many partial games.

Also, what are these symmetries you are referring to? Do you mean just simple geometric symmetries like reflections, rotations? The efficiency in analysis that would be achieved by accounting for equivalences under geometric symmetries is negligible. For example, even if we divide the number of positions by 8 (which is the maximum number of distinct positions obtained by doing rotations and reflections), that would still leave us with the same order of magnitude of distinct positions.

[towforce]

Of course, GMs know many more patterns. But we also know that GMs are nowhere close to beating even the current, less than perfect, top chess engines.

The engines are also doing deep searching: by comparison, the amount of searching a GM can do is so small that it's negligible in comparison with the engine search - so one could reasonably say that the human GM doesn't search.

You had mentioned the game of Nim in another post. Nim is different. It is an impartial game, and impartial games have a nice structure, as characterised by the Sprague-Grundy theorem. Such simple structure is not known for many partial games.

If the "simple structure" of Nim were not known, it could be found by searching for patterns.

This is the actual nature of much mathematical progress:

1. Spot a pattern

2. Tidy it up and formalise it into a formal mathematical theorem

Also, what are these symmetries you are referring to? Do you mean just simple geometric symmetries like reflections, rotations? The efficiency in analysis that would be achieved by accounting for equivalences under geometric symmetries is negligible. For example, even if we divide the number of positions by 8 (which is the maximum number of distinct positions obtained by doing rotations and reflections), that would still leave us with the same order of magnitude of distinct positions.

Let's take the symmetry of pairs of pieces as an example (there are many other types of symmetry), as this is what I think you're alluding to.

If the number of pieces is n, then the number of pairs of pieces is triangle(n - 1) (can also calculate it as c(n,2)).

So if there are 10 pieces, then there are 45 pairs of pieces, hence 45*8=360 symmetries just based on piece pairs alone.

Triangle number calculator - link.
Combinations calculator - link.

[Ras]

Draw death is on the way for computer chess. Once an engine becomes unbeatable, it can't really get any better.

And your idea is quitting chess altogether. That's like a cure that works by killing the patient. Hardly anyone cares about chess variants - it's not as if we hadn't had them for a long time.

[towforce]

Draw death is on the way for computer chess. Once an engine becomes unbeatable, it can't really get any better.

And your idea is quitting chess altogether. That's like a cure that works by killing the patient. Hardly anyone cares about chess variants - it's not as if we hadn't had them for a long time.

Human v human chess is not going to bigger boards: it would be exhausting for the players and not engaging enough for the viewers.

If too few people care about which chess engine is the best, then engine chess will not be going to bigger boards.

If, however, enough people care passionately about which engine is the best, then the bigger board's ability to magnify differences in chess knowledge could start to look attractive!

In terms of chess being the incubator for AI development, bigger boards could be an important step forward. At the very least, we should start considering 1-ply chess: let's start putting the "intelligence" into AI!

[Ras]

If, however, enough people care passionately about which engine is the best, then the bigger board's ability to magnify differences in chess knowledge could start to look attractive!

With a bigger board, the balance between short-range and long-range mobility (e.g. knights vs. bishops) will be different, and that difference will increase with board size. King mobility in the endgame likewise. It will still be a strategy game, but it won't be chess, and the knowledge required to succceed at that game is different from chess knowledge.

Which also makes it uninteresting for people to watch, because they would watch a game that they don't understand, and as I said, hardly anyone cares about chess variants, so there's no motivation for people to drop chess in favour of something else just for enjoying engine matches. These matches would be meaningless match result numbers.

Btw., the draw death was also talked about ten years ago. Then GPUs got powerful enough to run NNs, and then NNUE was invented. There are still positions that even NNUE engines don't quite understand, so there's still room for improvement. Improvement that will not happen if instead, we abandon chess.

[Uri Blass]

Draw death is on the way for computer chess. Once an engine becomes unbeatable, it can't really get any better.

And your idea is quitting chess altogether. That's like a cure that works by killing the patient. Hardly anyone cares about chess variants - it's not as if we hadn't had them for a long time.

Human v human chess is not going to bigger boards: it would be exhausting for the players and not engaging enough for the viewers.

If too few people care about which chess engine is the best, then engine chess will not be going to bigger boards.

If, however, enough people care passionately about which engine is the best, then the bigger board's ability to magnify differences in chess knowledge could start to look attractive!

In terms of chess being the incubator for AI development, bigger boards could be an important step forward. At the very least, we should start considering 1-ply chess: let's start putting the "intelligence" into AI!

bigger board is a different game and is not chess.

It is also not clear which engine is better with bigger boards because engine A may be better with 12*8 board when engine B may be better with 20*8 board.

[Uri Blass]

If, however, enough people care passionately about which engine is the best, then the bigger board's ability to magnify differences in chess knowledge could start to look attractive!

With a bigger board, the balance between short-range and long-range mobility (e.g. knights vs. bishops) will be different, and that difference will increase with board size. King mobility in the endgame likewise. It will still be a strategy game, but it won't be chess, and the knowledge required to succceed at that game is different from chess knowledge.

Which also makes it uninteresting for people to watch, because they would watch a game that they don't understand, and as I said, hardly anyone cares about chess variants, so there's no motivation for people to drop chess in favour of something else just for enjoying engine matches. These matches would be meaningless match result numbers.

Btw., the draw death was also talked about ten years ago. Then GPUs got powerful enough to run NNs, and then NNUE was invented. There are still positions that even NNUE engines don't quite understand, so there's still room for improvement. Improvement that will not happen if instead, we abandon chess.

I think chess knowledge is clearly relevant for bigger board.
The technique to win KRK is still the same and a lot of ideas are the same.

It is not chess but I disagree that people do not understand the game.

I expect top GM's to win in case there is a competition with big prize money with a bigger chess board.