search

stefanpeidli / GoNet

A students Project on GO
MIT License
6 stars 3 forks source link

Build Training Structure for Value Net #14

Open stefanpeidli opened 6 years ago

stefanpeidli commented 6 years ago

Please do the same as you did for the Policy Net for the Value Net (we won't save Board vectors here, just a triplet ~(black wins, draw, white wins), thus we won't need to use symmetry here).

Beno71 commented 6 years ago

Could you specifiy the task a bit more?

stefanpeidli commented 6 years ago

@Beno71 Build another TrainingData script, that loads sgf Games and saves the boards (+symmetries) this time not with a absolute "move played" distribution, but with a resultvector (amount of games that black wins in the end when this board appears, number of draws in the end, times that white wins). This is needed for the ValueNet. Is this clear enough?