i think similar approaches have been done. I can recall seeing it. Though in Backgammon they did train only by endresult and seemed to work fine. Originally anyway, now the have separate NN-for certain phases of the game
2016-01-30 18:07 GMT+02:00 Xavier Combelle <[email protected]>: > I had got an idea but I don't think I'm strong enough programmer to > implement it. (In particular I know quite nothing about neural network) > So I submit it here. > > If we have a neural network which is able to evaluate all the positions of > a board. > The following might help to improve it. > > From a position given: > check the max value of all the evaluations > go to the next level in the tree > check the min value of all the evaluations > if the min value < max value train the network at the root level to target > max value for the original position > else go to next level and continue > > The reason why I think it could help is because the evaluation at a deeper > level > should be easier and so better than at a less deep level. > > Please give all your feedback on this idea (even: it's a stupid idea/you > should implement it are welcome) > > _______________________________________________ > Computer-go mailing list > [email protected] > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://computer-go.org/mailman/listinfo/computer-go
