I declare that HCE is dead...

[Rebel]

When will search be declared dead?

Can search be put in a NN as well?

[connor_mcmonigle]

When will search be declared dead?

Can search be put in a NN as well?

It's unlikely that a neural network will replace search. However, move ordering heuristics as well as some LMR, pruning and extension decisions are likely to be replaced with something akin to a policy network in the immediate future in alpha beta search engines in my opinion. FWIW, HCE were effectively one layer networks anyways so the distinction is somewhat meaningless.

[towforce]

When will search be declared dead? Can search be put in a NN as well?

The problem with chess for AI study is that for decades search has been killing all other techniques.

The REAL answer to chess will be to solve it: then EVERYTHING ELSE (search, NN, HCE etc) will ALL be dead!

[Milos]

When will search be declared dead?

Can search be put in a NN as well?

That 100% equivalent question to "Can one encode 32-men TBs into NN?"
I guess you know the answer.

[Milos]

The REAL answer to chess will be to solve it: then EVERYTHING ELSE (search, NN, HCE etc) will ALL be dead!

We know how to do it for quite some time, i.e. we know the algorithm (Turing machine) that 100% provides an answer. We know how to generate 32-men TBs. We "just" don't have enough resources for this task .

[towforce]

The REAL answer to chess will be to solve it: then EVERYTHING ELSE (search, NN, HCE etc) will ALL be dead!

We know how to do it for quite some time, i.e. we know the algorithm (Turing machine) that 100% provides an answer. We know how to generate 32-men TBs. We "just" don't have enough resources for this task .

As discussed previously, I think it can be done with today's technology and without the 32 piece tablebase. If you want to discuss it again, probably should start a new thread, but, briefly, might be possible by recursively calculating what conditions must be true to win material (assuming that the king is "just" high value material), and discarding the outcome each time a generated condition is found to already exist in the condition list. The recursion finishes when no new conditions are found. Alternatively, might be possible by the following iterative process:

1. Get a set of positions and correct evaluations
2. Use curve fitting to find the simplest algorithm that provides the correct evaluation for all positions
3. Find a position that the algorithm evaluates incorrectly
4. Add it to the set of positions and correct evaluations
5. Repeat

This would require curve fitting in a large number of dimensions, for which existing software isn't great right now, but I think I can come up with a way of doing that.

[Rebel]

When will search be declared dead?

Can search be put in a NN as well?

That 100% equivalent question to "Can one encode 32-men TBs into NN?"
I guess you know the answer.

Not necessarily, there are different questions, such as, is it possible to train a NN for a safe number of reductions for each position.

[towforce]

Quick demonstration that you don't always have to check every permutation: the travelling salesman problem has been solved to 85900 towns. If every combination had been tested, that would be 85900!=9.6*10^386526 combinations. That didn't happen: instead, the overwhelming majority of combinations were eliminated mathematically without having to be individually checked.

We don't yet know how to do that for chess. However, as stated in a previous thread, for me the evidence strongly suggests that relatively small algorithms could evaluate most chess positions well (this is, of course, well short of saying that chess could easily be solved).

[Madeleine Birchfield]

Quick demonstration that you don't always have to check every permutation: the travelling salesman problem has been solved to 85900 towns. If every combination had been tested, that would be 85900!=9.6*10^386526 combinations. That didn't happen: instead, the overwhelming majority of combinations were eliminated mathematically without having to be individually checked.

We don't yet know how to do that for chess.

Yes, we do, it is called using minimax search and alpha-beta from the starting position instead of iterating through every position in a 32 piece tablebase. There are many positions with 32 pieces that are simply impossible to reach from the starting position, such as most chess960 starting positions, and many others that are a result of blunders, which would be eliminated in minimax and alpha-beta.

Minimax and alpha-beta could be used in any weighted directed graph, such as those found in the travelling salesman problem.

[towforce]

Quick demonstration that you don't always have to check every permutation: the travelling salesman problem has been solved to 85900 towns. If every combination had been tested, that would be 85900!=9.6*10^386526 combinations. That didn't happen: instead, the overwhelming majority of combinations were eliminated mathematically without having to be individually checked.

We don't yet know how to do that for chess.

Yes, we do, it is called using minimax search and alpha-beta from the starting position instead of iterating through every position in a 32 piece tablebase. There are many positions with 32 pieces that are simply impossible to reach from the starting position, such as most chess960 starting positions, and many others that are a result of blunders, which would be eliminated in minimax and alpha-beta.

Minimax and alpha-beta could be used in any weighted directed graph, such as those found in the travelling salesman problem.

Good point, Madeleine!

However, I don't think these tree pruning methods are likely to be enough to crunch out chess from the starting position (it is interesting, though, that in the 64 years of computer chess nobody has yet found a way to force the win of material from the starting position - and some programs are now searching quite deeply from it).

Also, the goal I have in mind is to be able to evaluate any position quickly and accurately: the "quick" part of that would preclude search: I have suggested it might be possible to generate a set of rules to say whether material can be won, or to create a smallish algorithm which produces the correct answer for all positions for which the correct value is known.