[computer-go] Improvement of UCT search algorithm

[sylvain.gelly at m4x.org]

Łukasz and Don,

we've try theses ideas on MoGo but without real success. We've not tried so 
long because it was at the beginning of MoGo's developpement, and MoGo was 
very weak. Perhaps now this kind of improvement would be more visible?

We also think that this kind of ideas could be an improvement to handle the 
non stationarity property, but this can also increase the impact of the noise 
(which is bad of course).

For the problem of how to test the parameters and ideas, I agree there is no 
good solution. For MoGo our main test is doing a lot (often 400) of games 
against gnugo 3.6 at default level with MoGo using 70000 simulations per 
game. The good think is that it is a real test, the drawbacks is that it 
takes a LOT of computation time, and we only improve MoGo against gnugo. 
There is not always transitivity. This fact can easly be seen on the results 
on cgos. We are comparatively quite strong against gnugo (and gnugo based 
programs), but quite weak against MC based programs.

If anyone have a good test, please let us know :-).

Sylvain