> Since then it has also become clear that Monte Carlo tree search > benefits from increased computing power a lot more than any previous > algoritms in GNU Go. The Monte Carlo code has its own board > implementation in engine/montecarlo.c and was designed with parallel > computing in mind, although there is no actual support for that in GNU > Go yet.
My impression of the MC engine is that it is still weak at life and death. I can often kill it though it plays very good yose. If it were stronger at life and death, it would be strong for 9x9. > There's no particular need to do anything drastic with the existing > code. The only thing that matters in the short term for making GNU Go > stronger is to catch up with the Monte Carlo development, especially > make the Monte Carlo code scale better to larger boards. Key ideas to > look into there are RAVE (rapid action value estimation), > progressive widening, and improved move ordering. Improved move ordering may be mainly a tuning matter. What is the process for tuning the existing Monte Carlo code? How can we see what the code is doing? I don't think these questions are fully answered with the existing doc. Dan _______________________________________________ gnugo-devel mailing list [email protected] http://lists.gnu.org/mailman/listinfo/gnugo-devel
