Can chess be analytically "solved" in the forseeable future?

[henryzz]

Originally Posted by Xyzzy

http://arstechnica.com/gadgets/2016/...-in-the-world/

Sounds like this sort of method could have a future in chess but still needs some work.

[axn]

Originally Posted by henryzz

Sounds like this sort of method could have a future in chess but still needs some work.

Giraffe chess engine

[Xyzzy]

http://arstechnica.com/information-t...human-players/

[Prime95]

Originally Posted by Xyzzy

http://arstechnica.com/information-t...human-players/

Could a DeepMind chess program beat Stockfish (or at least play superior chess such that it could win TCEC)? Does anyone have an opinion on whether Stockfish can be improved significantly other than searching deeper with faster computers?

Since four years have passed since this was last discussed, is the general consensus that perfect chess results in a draw? Does anyone here believe otherwise?

[chalsall]

Originally Posted by Prime95

Does anyone have an opinion on whether Stockfish can be improved significantly other than searching deeper with faster computers?

What is the storage space of all workable chess moves?

It seems a bit like a "rainbow attack" to me.

Go, on the other hand....

[retina]

Originally Posted by chalsall

What is the storage space of all workable chess moves?

Shannon number ...

[Dubslow]

Originally Posted by Prime95

Could a DeepMind chess program beat Stockfish (or at least play superior chess such that it could win TCEC)? Does anyone have an opinion on whether Stockfish can be improved significantly other than searching deeper with faster computers?

Since four years have passed since this was last discussed, is the general consensus that perfect chess results in a draw? Does anyone here believe otherwise?

It's.... possible, though it would take a lot of work. The engines use a *lot* of heuristics to guide/prune their branch searching, and while they are human-decided heuristics, a neural network would take a *lot* of training, in my decidedly amateur opinion, to surpass the human heuristics.

Importantly, the key difference is that the branching factor for each move is around 30-50 for chess versus 200+ for go, so chess is unlike go in that for go, neural networks make searching *at all* viable/feasible, giving them a massive advantage over humans/more brute force searches alike. Chess, on the other hand, has always been susceptible to a brute force type search, even if the human heuristics have come a long way in being smart about what remains fundamentally brute force.

Basically, it's possible but would require a ton of research/time/effort.

[chalsall]

Originally Posted by Dubslow

Basically, it's possible but would require a ton of research/time/effort.

Which Google has.

[Dubslow]

Originally Posted by chalsall

Which Google has.

But it wouldn't exactly be novel research, unlike for Go.

By the way, here's the paper published by Deep Mind a year ago, shortly before the match against Sedol.

https://github.com/brilee/MuGo/blob/...lphago2016.pdf

[a1call]

Originally Posted by Dubslow

Basically, it's possible but would require a ton of research/time/effort.

I think that if it can be done, then the best way to do it is by brute force on distributed computing. And I think money/complexity is/are not the issue(s), since the principal program can be written by anyone who has ever written a chess program and there are many of those. The only obstacle is the astronomical number of possible moves. I think one improvement (applicable to chess in particular) to the program would be to add a hierarchical aspect to the distributed computing. Say the top node assigns one possible move and its children to (up to) 32, 2nd level nodes, and so on down the hierarchy. On the surface and without doing any math, this seems very doable. But one should always remember the story of the inventor of the chess who asked for a grain of wheat doubled on each chess square would result in more grains than have ever existed on Earth.
I think "Black" is the absolute winner of chess, based on nothing more than I always play black,

[a1call]

Originally Posted by Cazzeo

As far as solving the game, memory is one of the biggest constraints. When you can make 5, 6, 7, 8, etc. moves in a row, the primary way to reduce the ply count is by recognizing transpositions (i.e. that Nf3-d4-c6-xe7 is the same as Nh3-f4-g6xe7), of which there are many. So you need giant transposition tables.

I assume transpositions are repeat configurations. I think one way to handle the large number of repeated configurations is to limit it to hierarchical nodes, which only analyze a limited depth of the parent configuration. There will be overlap with neighboring nodes, but distribution of the repetition analysis should break it down to manageable tasks per hierarchical node since you are looking for repetitions which are slightly different than your parent node and are not astronomically high in number.