Sometimes the top move gets 99% or even 99.9% of the playouts, even though it's not clearly the best move. It might actually be a big mistake, but better moves were unlucky in the few playouts they got and were never reconsidered.
I've found myself wishing for something similar to Ingo's proposal, but I would do it very simply on the UI level. If the Move List allowed you to select one row to mean "ignore moves above this", that would be enough. By default the top row would be selected and CrazyStone would do whatever it normally does. If the first move is getting too much attention, just click the second row and then CrazyStone will consider all moves except the first one. Another nice thing would be a column that shows the "main variation" for each move. Sometimes CrazyStone is concentrating in a pointless ko-threat and what you care about is the move after that ko-threat and the response. Changing between analysis mode and "resume game" is also very cumbersome if you're analyzing an ongoing game. When a new moves is made, you have to: 1) Resume 2) Make the move 3) Suspend 4) Analyze Only 2) should be enough. There's a long way to go to match chess engines, but CrazyStone is quite nifty already. :) Johannes On Sat, Nov 15, 2014 at 9:06 AM, "Ingo Althöfer" <[email protected]> wrote: > Hello all, > > CrazyStone is a fine (commercial) go program. > It has a nice analysis mode, where for a position > %-values for the top 20 or so moves are shown, > including the numbers of playouts for these moves. > > When analysing the games from the codecentric Challenge > very often the top move got more than 90 % of the > playouts. Such a distribution may make sense for > a playing program. However, for ANALYSIS mode it would > be better when SEVERAL candidate moves got many playouts. > > QUESTION: Who has ideas how such a k-best mode for > MCTS might best be organized? (Here k is the number > of candidates who get large numbers of playouts - and thus > "precise" evaluations.) > > Ingo. > > PS. In computer chess such k-best modes are THE tool > for analysis (in GM practice, in correspondence chess ...) > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
