Is the true scalability of mogo (against a variety of programs, or against people) less than the rating study indicates?
Unfortunately, I think that the answer is yes (very very). In particular the tuning has been performed on small games, and I guess the new very fast (parallel) versions of mogo would be better with - more exploration; - more diversity in the MC part. In particular, some of the last improvements of mogo were negligible (if positive) for small numbers of simulations, but are incredibly efficient for large numbers of simulations. It's not very surprising: if you have more computational power, you have to explore more, and you can afford a bigger diversity in the MC-analysis. The parallel version has a huge speed-up, which has also an effect on the ranking vs humans, but the effect is much lower. What we need now is probably much more a qualitative change in the MC, whenever this qualitative change is computationally expensive. Thanks to clusters, we can afford expensive Monte-Carlo. In my humble opinion, if in the same code we have: - the mogo parallelization (30 lines of MPI, 30 lines of C++, not more...) - very good MC simulations (e.g. the ones with the best results for the "tsumego") - a big cluster then you have a fantastic Go program. Olivier _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
