[x] Phase 1: Take a single hand as inputs, label is the nth of tile to discard.
[ ] Phase 2: Take a single hand and the rest of all the observable information as inputs, label is the nth tile to discard.
[ ] Phase 3: Take a single hand and the rest of all the observable information as inputs, label is the nth tile to discard and the action player made(chi, pong, kang, riichi, ron).
[ ] Optional networks: guess the probability of one/multiple opponent(s) hand information
After Riichi, the tile that the opponent(s) are listening to.
Probability of all the tiles.
Alternative to simple neural network
[APPLIED] CNN for a 'convolution' of input features.
RNN/LSTM for a vector input of all the tiles(as the amount of the observable tiles is changed per player move).
Neural Network Structure
Alternative to simple neural network
Resources
Downloading Records Records Decoder