CNN features
I intend to learn a Convolutional Neural Network (CNN) for the AI. For this I need a game encoder.
The data samples will be tensor of dimension (92 * N * (6 + T + C)), where N is the number of players, T is the number of tricks, and C 0 if there is no call, 1 otherwise. The dimensions are:
- The event data (8 possible event types + the slam declaration + 5 possible bids + 78 possible cards = 92 dimensions);
- The player;
- The steps and later tricks: one page for the deal, one
Just to say here in passing, filters should maybe be of dimension 14 * N * N, in order to capture directly the relationship between cards of the same suit (there are 14 of them), and to capture a trick. Of course, the filter at the second dimension should be cyclic.
So, make a program to extract those CNN features (tarot-cnn-features). The program prints a line for each decision (i.e. a line of 92 * N * (6 + T + C) numbers for each time a player whom we know the cards play). The score is not a problem, because we already have tarot-status score.