Hi 1. Just a small precision first:
Ingo Althöfer wrote: > Terry McIntyre wrote: > > ... My pet peeve is the KGS "score estimator", which is often wildly > wrong > As explained by others a "strong SE for ALL positions" is equivalent > to a strong program. This is only true if you replace the word “strong” by “perfect”. It is trivial, that a perfect evaluation function gives a perfect player using a 1-ply search, but that is not true for strong evaluation functions. Converting a strong evaluation function into a strong program is still a hard problem. “Almost” good moves are frequently worse than random moves. (E.g. a move that “almost saves” a group that cannot be saved or “almost kills” a group that cannot be killed gives one prisoner to the opponent and wastes one turn. This is worse than pass while random moves are with a high probability better than pass.) 2. Now my proposal: An MC based SE (Score Estimator). Use a strong MCTS program unmodified, except in a small detail to collect data about by how many points each simulation is won. Play a feasible number of simulations like 20K. Build the distribution of the result (= The histogram counting how many wins were found by each result). Otherwise, let the program follow the tree as it does unmodified. In this distribution find the 3 quartiles (the komi shift that produces Q1 25%-75%, Q2 50%-50% and Q3 75%-25% wins for each). If IQR (Interquartile range) = Q3-Q1 > some threshold (say 25 points) output: Q2 as “B+3.5 with large error margin” else output: Q2 + IQR as “B+3.5 IQR(W+1.5, B+11.5)” This would be a state of the art SE and the number of simulations could also be a parameter controlled by the user to adapt the precision to his hardware and patience. After all, it is easier to get a SE from a strong program than the other way round ;-) Jacques. _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/