Hi I have some math questions about MC and would appreciate some link to stat course (i should have been more attentive at university ;)
- the original MC used mean_score/standard_deviation as criterion I understand why it is a good thing. I found that std/mean is internal coherency factor Is this equivalent to choosing the winning probability? (i guess no, but i have doubts) - Instead of score i tried using wining probability (score=1 or -1) With a fixed number of simulation, standard deviation is no use as we can calculate it from the mean probability and N_simul. So i need higher order estimator to check how many simulations are needed at a given game_status for a reliable estimation (for something like incremental alpha-beta: make a first pass, cut branches, redo si I don't remember how to compute this without doing it manually : Split the N_simul in M group or Ns=N_simul/M Run the M group and for each compute mean Then have the real mean, and the std of groups This processus can stop before reaching M, if we find that the std between groups is reasonably small This for something like progressive alpha-beta: - make a first pass with Ns, cut rotten branches, do Ns more simulation, cut branches Alain _______________________________________________ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
