Run analysis_main.py for the main results with PNNS using JSD.
Uncomment the game you want to analyze (the script will only run with the first game in the list); change n to the number of runs (different train/test splits)
Run any_model_test.py to test JSON-Bag with any other machine learning model. Here, the JSON-Bags are vectorized. Choose the game same as above. Choose model the same way; any classification model can be plugged in, assuming the same API as sklearn.
The tokenizer is the function tokenize(...) in tokenizer.py that takes in a JSON object (loaded as dict), the tokenization mode (ordered, unordered, both, char, more details in the paper), and the option for binning/pairing x, y coordinates in grid-based games, and output a list of tokens.
| Game | Tokenization mode |
|---|---|
| 7 Wonders Dominion Sea Salt and Paper Can't Stop Connect4 Dots and Boxes |
unordered unordered unordered both ordered ordered |
TODO: Implement the tokenizer as a class with more robust customization
All games have these features:
- Game tick (engine specifics)
- Turn count
- each player's score at game over
- rate of change and intercept of each player's scores throughout the game (except for Connect4)
- Scores of a player are recorded periodically throughout a game trajectory into a score vector
$\mathbf{s}$ , a linear regression model is fitted to predict$s_i$ :$w \times i + b = s_i$ , where$i$ is the index of the score. We extract$w$ and$b$ for each player as features.
- Scores of a player are recorded periodically throughout a game trajectory into a score vector
In the folder const, each game has a gamename_const.py with a list features containing game-specific handcrafted features.
The games used for the experiments are implemented in TAG Framework. The fork used by the experiments is https://github.com/dienn1/JSONBagTAGFork. All data is generated with this fork. The specific parameters used for the playing agents can also be found there.
Due to size, the raw data is omitted here. In each game folder, there are some example raw game trajectories (concatenation of all JSON game states in a trajectory). The folders also include the full JSON-Bags of game trajectories