A neural network for predicting best dice to keep in a given throw in Ten Thousand dice game.

I have an open question about this code at DataScience on Stack Exchange

Tensorflow model for predicting which dice to keep for a given roll. Thus far it works well for predicting 1's and 5's, but not three of a kind, three pairs or straights. I may need to use a different optimizer, or add more examples of special scoring situations to the training set. I'll ask at Stack Exchange Data Science.