To show quickly how ubiquitous the appearance of several extrema is in more dimensions, I am using a simple example in 2 dimensions. The theoretical function is separable C(X,Y)=A(X)*B(Y), with each A and B having only one extremum, and being pretty smooth. This, I think, is a realistic interplay for chess parameter global tuning. The resulting C(X,Y) is hereLaskos wrote:In several dimensions, are you doing the multi-dimensional quadratic regression? I don't read the CLOP documentation, but if it's doing this, the result will almost certainly settle on the broadest local extremum, not on the global one. In many dimensions there will almost always be a hell lot of local extrema, even if each 1-dimensional function is very smooth. UCT is not exactly what I meant, it scales badly with the number of dimensions, rather on something like VEGAS by Lepage, adaptive recursive method where the data points are concentrated in the regions close to exrtrema. It scales extremely well as the number of dimensions go, linearly or a power close to 1, if I remember well.Rémi Coulom wrote:Yes, CLOP would probably miss the global optimum with such a function.Laskos wrote:To exemplify quickly, we have a theoretical curve to play with for finding a global maximum:
But I don't worry about it, because I don't expect such function shapes will occur when tuning parameters of a game-playing program.
An algorithm such as UCT would find the global optimum. But such global optimization methods don't scale well to more than a few dimensions.
Rémi
Kai
It has one sharp global peak and a broad local peak. Adding white Gaussian noise, this gives
and performing two dimensional quadratic regression on this noisy data, the fitted function is
The peak reproduced is the broad local one, not the global one. This is just one example when from an apparently innocuous separable function into some smooth functions, one gets this situation. In many dimensions these problems are almost unavoidable.
I don't know if that is what CLOP is doing, but anyway, a brief example of what could happen.
Kai