I am trying to write self-play RL (NN + MCTS http://web.stanford.edu/~surag/posts/alphazero.html) to "solve" a board game. However, I got stuck in designing boardgame same (input layer for NN).

1) What would be the best way to represent each cell, if there are ~10-100 cells in a game which could be occupied by any playing card. Do one-hot-encoding and get 52 nodes for single cell ([0, 0, 1, ..., 0]) or just divide card_id by total number of cards and get single node for each cell ([0.0576...])?

2) Is it a good/bad practice to help NN by adding additional input that could be derived from the other nodes? For instance, imagine the game where whoever has most red cards wins. Input is 10 cards, and I am adding new input node (number of red cards) to emphasize it. Would that lead to a positive result or doing something like that is bad?

3) Would it help to reduce the number of illegal moves and increase the performance of NN by creating additional input stating which cards are available now and which are not?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.