5 Machine learning with tarot

There are support for different paradigms of machine learning in tarot. The library provides a set of hand-crafted high-level features, and tensor-shaped features for use with a Convolutional Neural Network.

The high level features that way: for each game to learn, we can extract some game features. The features are, for instance, how many trumps we have, whether a partner may still play, and so on.

Each move for each player implements one or multiple strategies. For instance, one strategy is “play the called suit”, or “take the trick”, or how many points you play. This shows that the strategy may not be binary.

Then, we record the final score for each player. So, for a given strategy, we learn how many points we might get in the given game state. It is important to consider the interactions between features: for instance, we should not have the same model for all game variants.

Please note that these functions are tailored for machine learning: they are redundent with the rest of the API, and they never fail, but the value may be meaningless. Both features and strategies are only meaningful if the next player is the main player of the given game (i.e. we know her cards), and there is uncertainty (we don’t know the cards of the other players). If we know, the learning will not be very useful because the data set will not be taken from the same distribution as the test set.

On the other hand, the library also provides support for tensor-shaped data.

• The individual pages encode game steps. The first page is for the deal, the second page for the bids, the third is for the slam declarations, then there is an optional page for the call (if the game variant allows it), then a page for the dog (possibly filled with zeros), likewise for the next page for the discard, then a page for all handfuls, and then a page for each trick. In order to have the same number of pages for all variants, the remaining pages are padded with zeros until there are 32 pages.
• Each page contains a line for each player. The first line is for player number 0, not for the first player in the trick. No line is allocated for the dog. In order to have the same number of rows for each game variant, the rows are repated until there are 6 rows.
• Each page contains a number of columns. The first 8 columns encode the event type as one-hot: all columns are set to 0, except the column encoding the actual type which is set to 1. The types are in this order: deal, bid, declaration, call, dog, discard, handful, card. The type is encoded on each relevant line, i.e. on the line corresponding to the player adding that event. So for instance, on the call page, only the call event column is set, and only for the taker. The next column is for the slam declaration: 0 for no slam or not decided yet (it is possible to know which applies, by looking in the ’declaration’ type column: if it is set to 1, then no slam is declared, otherwise the player has not decided yet). Then, 5 columns encode the bid: pass, take, push, straight or double keep. Finally, the 78 other columns encode the card. In the case that the event involves a set of cards, each corresponding card is set at 1 for the relevant player (the player who has been dealt the cards, the player who has shown the handful, the taker for the dog and discard (if known), …). So, there are 8 + 1 + 5 + 78 = 92 column.