Passer evaluation

[lithander]

A single layer network without a activation function is mathematically identical to Piece-Square Tables
More layers have deeper terms and would be easiest to think of looking a few plies ahead.

The deeper the network the more training data you would need. If the generation of enough annotated positions was not practical for tuning PSTs or equally simple eval terms than this rules out the NN approach. Afaik the top chess engines already use billions of positions as training data and it wont help that this variant is played on a 12x12 board.

[dangi12012]

The deeper the network the more training data you would need. If the generation of enough annotated positions was not practical for tuning PSTs or equally simple eval terms than this rules out the NN approach.

The beatuy of NN learning is that the single scalar error value for example (-20.0f) multiplied by the scaled inverse matrices per layer - will update all weights in all matrices at once accordingly. So deeper networks dont need more training data in theory. In practice only a few more iterations because of compounding error.

The keyword here is backpropagation and a 8x8x(types of pieces = 13) neural network is really really really small compared to a typical image NN. Going to 12x12 is laughibly small when compared to image algos that also need to look at millions of images (ie positions) but have input sizes that are 224×224 for example. So 12x12 chess is 348 times smaller than image recognition. Not to mention that for chess an input is not float rgb but only the piece type or zero for empty.

Most websites dont do machine learning justice and it is explained in convoluted terms while its far simpler than handwritten heuristics. The math fits on half of a napkin.