Chess engines can soon simulate tablebases for any number of pieces?

[andytoh]

With engines and computing power so strong now, can we assume that one day not too far in the future that chess engines, if we allow them ample time to compute, can "simulate" a tablebase for any number of pieces? For example, take any position (even possibly the starting position), let the engine compute for days or even weeks. If the evaluation is between -0.2 to 0.2, then we can say with 99.9% certainty that the position is a tablebase draw? The question that we will naturally wonder is what (ultimate) evaluation would we claim that the position is tablebase win for one side. If the evaulation is more than 5.0, I think we can safely say that the position is a tablebase win for White, though the number of moves to checkmate will not be computable (that is where the engines fail to match the tablebases). Thouugh we may never know if an evaluation of +1.5 is a tablebase win for White or still a draw (with very accurately play by Black), we can at least know with great certainty that a given position is a tablebase draw, if we let these future engines compute long enough (e.g. the evaluation never exceeds +0.2 after one month of computing). Your thoughts on this hope for a simulation of the ultimate tablebase solution to chess?

[koistinen]

To simulate any number of pieces, try limiting hash table size and see how engines handle a small number of pieces.
The ones it gets wrong when searching for a few minutes, try searching a few hours and see how often that fixes the problem.
I think it won't work for all positions, but seeing how often the engine gets w/d/l right without a tablebase would be an interesting statistic.
As the engines are developed using the knowledge gleaned from tablebases, expect them to perform worse for positions where no tablebases have been computed.

[h.g.muller]

Determining the W/D/L of a given position with a 99.9% probability to be correct is not 'simulating a tablebase'. Or with 99.999% probability. For one, an EGT will offer 100% correctness, and thus has an infinitely smaller probability to be wrong. But the most important distinction is of course that an EGT offer this for all positions with the given material, not just for one, and that it will offer those by a lookup taking microseconds, rather than weeks of computation.

Engine evaluations are pretty meaningless. I have seen Stockfish and Houdini evaluate a totally lost position as +8. Now that engines are using neural nets for evaluation they have become even less reliable.