[Computer-go] Parameter Optimization with Local Quadratic Regression

Rémi Coulom Remi.Coulom at free.fr
Wed Mar 17 08:45:07 PDT 2010


I have recently put my recent results about parameter optimization on 
that web page:

With source code and slides of a presentation:

Summary: In this presentation, I'll talk about black-box optimization
from binary response. The motivation for this research is the
optimization of parameters of a Go-playing program from the observation
of game outcomes (win or loss). For this application it is safe to
assume that the function to be optimized (the winning rate) is a smooth
function of parameters, and it has no tricky local optima. The main
difficulty is dealing with noise. The first method that was tried is
UCT. But it approaches optimality like 1/sqrt(n), which is very slow.
Local quadratic regression performs much better. Empirical measurements
confirmed an intuitive proof that it should approach optimality like

Slides of the presentation might be difficult to understand without 
explanations, so don't hesitate to ask questions.

I am also sending this message to test the new mailing. I am not 
receiving any message at my univ-lille3 mailbox. I also receive no reply 
from messages sent to Computer-go-request at dvandva.org. I have also had 
reports from other former members who don't receive any message and who 
thought the list was still out of order. I figured this out from private 
communications with other former members. It may be that other former 
members are also in the dark. It would be very good to have a working 
web interface to this list again.


More information about the Computer-go mailing list