Evaluation is BY DEFINITION what you do in the leaves of a tree search where the game has not yet ended by rule (such as mate), to heuristically calculate a number that expresses the relative desirability of the position (assuming the goal of both players is to win). A unique (i.e. single-valued) mapping into the set of numbers is the mathematical definition of the concept 'function'. The unique value delivered by the NN's 'value head' in response to the board position and a few predecessors (together forming the current game state) thus is an evaluation function.

One would think even an idiot could understand this...

BTW, (and not that this matters...), the terms in AZ's NN are also hand-crafted: the topology of the NN (number of layers; which cells to connect with which, what non-linear transformation to apply to the sum of inputs in each layer) was consciously designed. Only difference is that this design did not take any game-specific knowledge into account, but is 'brute force', in the sence that you use such an overwhelmingly large number of terms that those terms that would be useful for a particular game are amongst those.

Any conventional evaluation can in fact be implemented as a neural net, typically with some 10-1000 neurons. But if you want such small networks to be any good you should indeed very carefully design the topology of their connections, to match the concepts that are relevant for the game. But throw in a couple of million neurons, interconnected as a multi-layer stack with an enormously high connectivity, and each topology you could desire will be a subset of that. Which can be 'instantiated' (through training) by giving large weights to the desired connections, and near-zero weights to the connections with cells used for other, independent purposes

chrisw wrote: ↑Sun Feb 24, 2019 9:37 am

Ed is correct when he says the output is a probability. AZ gives a probability measure of how many times this situation worked out as opposed to didn't work out, and for its policy a measure of how many times this move got tried as opposed to didn't.

Yes, it is a probability, so what? No one ever claimed anything else. It doesn't mater how you distort the numerical scale in which you express the desirability of the position, as long as you preserve the order. The rest of your claim is pure nonsense; in virtually all cases AZ would have never seen that particular position before, so how many times it worked out or not, or which moves could try would all be at a big fat zero. The probability it outputs is just an expectation on how well it thinks this position is going to do, based on features of it (such as material balance, pawn structure, king safety) it recognizes, and was tuned to guage their importance. Just as conventional evaluations do.