E is basically a measure of how well the evaluation function is able to predict the game outcome. I don't think the actual value of E is very interesting, because there are at least three factors that contribute to E:cdani wrote:All is working nicely, wining strength after each iteration on all the parameters. Thanks to all!!
I have used mutithread to do the computations, because it was very slow.
Current e = 0.0796898969 (28.2294%)
Has someone computed the "e" value for example for Stockfish? It makes any sense compute it? I really don't understand well the formulas, nor why this works. If someone tries to explain it, please, don't use maths :-)
Particularly, what is the error "e"? Where is the "error", mistake, think that is done badly? :-)
1. Mistakes in the game continuation causing the game outcome to not be equal to the game theoretical value of the evaluated position.
2. Tactical aspects of a position that can not reasonably be modeled by the evaluation function.
3. Bad values of evaluation function parameters.
The goal of the algorithm is to improve 3, without being mislead by the "noise" introduced by 1 and 2.
If there are lots of positions in the training data, the effects of 1 and 2 will on average not depend much on the parameter values, so by varying the parameters to reduce E, the parameter values get better. However, the individual contributions from 1, 2 and 3 to E is not known, so you can't say how good the evaluation function is based on the E value alone.