Brian Sheppard wrote:
Fuego uses a lower weight for distant moves than for nearby moves. I suspect that isn't much better than using uniform weight. I am hope that Martin or Markus will comment.
I measured a winning rate of 55.1(+-0.8)% of Fuego with weighted RAVE updates vs. the version with uniform updates. The experiment was done on 9x9 with 10K simulations. There was also a net-positive effect on the number of passes in the regression tests at the time. I did run a few tests against other opponents (GNU Go and MoGo) with longer time settings and on 19x19, but I didn't play enough games for a statistically significant result (it didn't seem to perform worse).
It would still be easy to investigate that deeper, if someone has the time and resources to run more experiments. The weighting of the RAVE updates is an optional feature in Fuego's UCT search and can be disabled with the GTP command "uct_param_search weight_rave_updates 0".
To be beneficial, it was crucial to make the weight a function of the relative move distance (w.r.t. the length of the simulation), not the absolute move distance. The exact formula is implemented in SgUctSearch::UpdateRaveValues(). It is scaled such that the average weight is still 1.
Markus _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
