The Next Big Thing in Computer Chess?

[towforce]

The next big thing in computer chess: some of you will already know my answer to this - find the deep underlying patterns in chess that determine whether a position is won or not.

In past discussions, I have given reasons for my belief that:

1. NNs tend to find a large number of surface (shallow) patterns rather than the deep underlying patterns

2. Deep underlying patterns are very likely to exist in chess

If deep underlying patterns can be found, I think it will become possible to play near perfect chess on cheap computers. Also, the method used to find the deep patterns could then be applied to other areas, meaning better AI at lower computing cost in many fields.

[CornfedForever]

The next big thing in computer chess: some of you will already know my answer to this - find the deep underlying patterns in chess that determine whether a position is won or not.

In past discussions, I have given reasons for my belief that:

1. NNs tend to find a large number of surface (shallow) patterns rather than the deep underlying patterns

2. Deep underlying patterns are very likely to exist in chess

If deep underlying patterns can be found, I think it will become possible to play near perfect chess on cheap computers. Also, the method used to find the deep patterns could then be applied to other areas, meaning better AI at lower computing cost in many fields.

That's 2 'if's':
1. If Bigfoot exists
2. If Bigfoot can be found.
Then...

Exactly what kind of 'deep underlying patterns' are you conjecturing about (#1) in the first place. Sorry...may have missed these previous posts you mention.

[smatovic]

The next big thing is a neural net driven search.

Isn't that what LC0 already does?

Yeah, but not Stockfish..,.

--
Srdja

[smatovic]

The next big thing will be when the GPUs and CPUs transparently share memory resources so that we do not have to copy to and from GPU memory.
Suddenly, engines like LC0 will become unbeatable.

It's not just the copy time that we save, it is a whole new programming paradigm.

You mean SoC?

I think SoCs are an astonishing piece of technology - but their strengths lie in their size (tiny) and their power consumption (milliamps). Having said that, one thing that Seymour Cray (famous supercomputer engineer and proprietor) used to emphasise strongly was reducing the distance between components - and not much does that better than an SoC.

Just for the files, the Apple M-series is considered a SoC with RAM chips mounted together in a system-in-a-package design...
https://en.wikipedia.org/wiki/Apple_M2

--
Srdja

[smatovic]

The next big thing in computer chess: some of you will already know my answer to this - find the deep underlying patterns in chess that determine whether a position is won or not.

In past discussions, I have given reasons for my belief that:

1. NNs tend to find a large number of surface (shallow) patterns rather than the deep underlying patterns

2. Deep underlying patterns are very likely to exist in chess

If deep underlying patterns can be found, I think it will become possible to play near perfect chess on cheap computers. Also, the method used to find the deep patterns could then be applied to other areas, meaning better AI at lower computing cost in many fields.

That's 2 'if's':
1. If Bigfoot exists
2. If Bigfoot can be found.
Then...

Exactly what kind of 'deep underlying patterns' are you conjecturing about (#1) in the first place. Sorry...may have missed these previous posts you mention.

God's Algorithm for Chess?
https://talkchess.com/forum3/viewtopic. ... 92#p932854
https://talkchess.com/forum3/viewtopic. ... 10#p932969

I like it.

--
Srdja

[smatovic]

The next big thing in computer chess: some of you will already know my answer to this - find the deep underlying patterns in chess that determine whether a position is won or not.

In past discussions, I have given reasons for my belief that:

1. NNs tend to find a large number of surface (shallow) patterns rather than the deep underlying patterns

2. Deep underlying patterns are very likely to exist in chess

If deep underlying patterns can be found, I think it will become possible to play near perfect chess on cheap computers. Also, the method used to find the deep patterns could then be applied to other areas, meaning better AI at lower computing cost in many fields.

Tactics playout vs. position evaluation, deep pattern vs. shallow pattern?

An AB NNUE engine does tactics playouts via search+qs and performs a position evaluation via (NN) heuristics.

The NNUE neural networks are not trained by simple shallow patterns, e.g. you take a chess position, run a depth n AB seach+quiescence-search and then label the position with a score and then use billions of those for training the NNUE network. There are also multilabel approaches. Hence I assume that NNUE networks encode both, tactics playout and positional play.

IMO it should be in theory possible to extract both into a classic knowledge base system contrary to nowadays neural network weights, would be an interesting project, extract chess knowledge pattern into a knowledge graph.

--
Srdja

[towforce]

The next big thing in computer chess: some of you will already know my answer to this - find the deep underlying patterns in chess that determine whether a position is won or not.

In past discussions, I have given reasons for my belief that:

1. NNs tend to find a large number of surface (shallow) patterns rather than the deep underlying patterns

2. Deep underlying patterns are very likely to exist in chess

If deep underlying patterns can be found, I think it will become possible to play near perfect chess on cheap computers. Also, the method used to find the deep patterns could then be applied to other areas, meaning better AI at lower computing cost in many fields.

Tactics playout vs. position evaluation, deep pattern vs. shallow pattern?

An AB NNUE engine does tactics playouts via search+qs and performs a position evaluation via (NN) heuristics.

The NNUE neural networks are not trained by simple shallow patterns, e.g. you take a chess position, run a depth n AB seach+quiescence-search and then label the position with a score and then use billions of those for training the NNUE network. There are also multilabel approaches. Hence I assume that NNUE networks encode both, tactics playout and positional play.

IMO it should be in theory possible to extract both into a classic knowledge base system contrary to nowadays neural network weights, would be an interesting project, extract chess knowledge pattern into a knowledge graph.

The best evidence against NNs encoding deep patterns is their relative weakness in end games: with all that knowledge applied to a simple position, that's where they should be stronger!

Of course, the art of the end game is "knowing what's going to happen well ahead", and hence to know what's important in the position now: somehow, the human GMs do this, but the ANNs don't.

There are other bits and pieces of evidence, like NN's needing to generate more positions than human GMs to play to the same standard. Although NN computer chess easily beats human GMs, the evidence is that the human GMs "understand" most positions better - which strongly implies that they have deeper patterns encoded.

Regarding a knowledge graph: it's basically a good idea. How are you going to find the best vertex on this graph, though? To put it another way, if you're at a given vertex, how are you going to decide which edge to travel along to get to the next vertex?

In 1997 I wrote an article for Selective Search (a British computer chess magazine) suggesting a way to write a chess program that would have an EF that was well suited the position: have a database of EFs, and use the indexes in the database that relate to the position that would pick the best available EF. These indexes would be related to things that are important in a chess position.

If that had been built, it would have been selecting EFs using chess position factors as database indexes: hence it would have been looking for the global optimal EF, not the local optimum that a knowledge graph might find.

[hgm]

I still believe it should be possible to program a fast algorithm for tactical analysis other than (quiescence) search. Somewhat similar to how N.E.G. selects its move, but taking into account soft-pins, overloaded pieces, subversion of protection... And that this algorithm would then either come up with a very good approximation of the QS score, or the admission that the position is too complex for a reliable evaluation. In which case it would indicate the move(s) for which it is uncertain of the outcome, which then could be searched.

BTW, using a 'policy network' in an AB engine should not be very difficult either. I think the main reason AlphaZero preferred to use MCTS rather than AB was that in the latter case it would be less obvious how to train such a network. But I guess that when you use a minimax search in the training, and then analyze the tree to order all moves that would have been able to produce a beta cutoff by the number of nodes it would have taken to search them through alpha-beta, you could train a NN with that info.

[smatovic]

[...]
Regarding a knowledge graph: it's basically a good idea. How are you going to find the best vertex on this graph, though? To put it another way, if you're at a given vertex, how are you going to decide which edge to travel along to get to the next vertex?
[...]

I am yet not sure, but I have it on my project list, "Theta", issue is RAM, would need a lot of it, and I have the idea to use RDF/SPARQL as graph database, project is planned for some distant future...
https://en.wikipedia.org/wiki/Resource_ ... _Framework

--
Srdja

[towforce]

This is just the bare bones of an idea - but how about...

* the computer shows you a chess game

* you rate how much fun it was

* based on that, the computer shows you another game

* you rate it again

* this continues until the computer is consistently showing you games that you love

* the computer uses this knowledge to play against you in a way that you really enjoy tremendously!



There might be some marriage advice hidden in there...