|
BGonline.org Forums
AlphaGo lessons
Posted By: Casper Van der Tak In Response To: Winning PR vs Losing PR--one thing impossible to measure (Stick)
Date: Wednesday, 30 March 2016, at 8:57 p.m.
It is possible to train a NN that will well predict what plays a human player would make (say open division strength), with likelihoods attached. Then you can use another program (or another part of the program) to identify the best play and the equity losses from worse plays. Then you can calculate the expected error for each roll/cube decision and have a measure for how difficult the plays you got to look at were, considering both the likelihood and size on an error. And you could link this to the number of decisions to get something in the same metric as PR.
I provided the concept so that is half the work, someone else can implement it :-)
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.