The neural network of the Stockfish chess engine

Re:The neural network of the Stockfish chess engine by kevincox on Wednesday January 13, 2021 @08:40AM (#25761819)

You don't need to consider every state individually to solve a game. For example if you have a game on a 100x100 board and the only goal is to move your piece to the other side you can know you have found the optimal solution if you can do it in 99 moves (assuming that you move 1 per turn).

In the chess scenario if you can prove that there is a set of moves where you win no matter what the other colour does then you have solved the game. You don't need to consider every possible game, just the games reachable within this set of moves.

Solving chess would be proving that "From starting position black can win every time" (or the same for white). You don't need to prove how to win from every possible board state.

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by danielbarla on Wednesday January 13, 2021 @10:36AM (#25763270)

I agree, and kind of included this in the "so, unless you had some way of dramatically compressing or simplifying this search space..." disclaimer.

Having said this, the nature of many chess endgames suggests that such a proof is not really possible, or at least would not be "simple". As an example, tempo / opposition flips games from draws to wins, etc.

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by pretendscholar on Wednesday January 13, 2021 @09:54AM (#25762677)

My money is on perfect play drawing every time.

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by kevincox on Wednesday January 13, 2021 @10:27AM (#25763128)

It would be interesting to take bets on the solution :)

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by wizzwizz4 on Wednesday January 13, 2021 @09:56AM (#25762710)

But you won't find a set of moves. You have to find a tree of moves, which is substantially larger.

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by kevincox on Wednesday January 13, 2021 @10:28AM (#25763141)

This is true. It is still a massive problem space but my point is that you don't need to consider every possible game or state.

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by wizzwizz4 on Wednesday January 13, 2021 @10:40AM (#25763328)

Not to prove it always-winning; you only have to go down every possible branch for the other player for that (i.e. √ of the number of states you'd otherwise need). To find it, though? Well, unless you hit lucky, you're going to be considering pretty much every possible game or state, except where you can find shortcuts.

[ Parent | on HN ]

Re:The neural network of the Stockfish chess engine by kevincox on Wednesday January 13, 2021 @11:00AM (#25763706)

It is a bit more than that as it is unlikely that the always win sequence would be a fixed list of moves. You would need to show that for you each move the other player makes you have a reaction that also wins.

[ Parent | on HN ]