0

There is a two-player game (discrete, deterministic, perfect information and so on) where - in some but not all states - a few moves may be equally good; i.e. they are symmetric and expert player will expect the same outcome from any of those moves. How to train neural network agent to treat those moves equally and how to measure accuracy for such approach?

Or maybe such approach doesn't make sense at all? Maybe policy should stick with only one variant (i.e. "go left side") so we can point only one move as designated valid one for given game state?

tomash
  • 101
  • 1
  • 1
    You need to give us more details. What exactly is your data? What is the model? Usually any model learns from some data, so if both scenarios lead to same outcomes, they are "considered" as same by the algorithm. – Tim Apr 09 '19 at 14:19
  • @Tim I left details out on purpose, but in current approach I have sequential Keras neural network, mostly binary-encoded game state and one-hot encoded move at output (~125 possible moves, half or less valid for given state, a dozen or so reasonable). Not sure how to implement state-agnostic "equally good move" output encoding (is it even valid approach?) – tomash Apr 09 '19 at 14:32
  • How is the information that the move is "good" passed to the network? What are you predicting? Please edit the question to give us more info. – Tim Apr 09 '19 at 14:38

0 Answers0