Game of Life Emulator

An experiment in game emulation through machine learning. The goal was to emulate a game using a machine learning model in the simplest way possible.

At each time step, the state of the game grid is fed into a "cell predictor" which predicts the new state of the grid. The game grid is then updated to this new state.

The cell predictor can be configured to be either a hard-coded one which explicitly follows the usual Game of Life rules, or a machine learning-based one which draws its predictions from an ML model. By default, the ML-based predictor is used.

Running the game

Ensure you have Python (3.10) installed and a virtual environment created and activated. Install dependencies with:

pip install -r requirements.txt

You can run the game with:

python pygame_life.py

Command-line help is available with:

python pygame_life.py --help

Command-line options

--help: Shows command-line help.

--predictor: The predictor is the component which controls grid updates. Default value: model.

• model is a machine learning model-based predictor.
• explicit is a predictor with the Game of Life update rule explicitly coded in.

--update-mode: The grid update mode determines how the predictor's output update the grid. Default value: normal.

• normal means the outputs are thresholded to 0 or 1 and the thresholded values are written to the grid.
• probabilistic means the outputs are interpreted as probabilities p, where the cell becomes 1 with probability p and 0 with probability (1-p).
• soft means the predictor outputs are written directly to the grid. The predictor outputs are fractional values, so each cell is somewhere between "alive" and "dead".

Retraining the model

The ML model is a fully-convolutional neural network trained using train_cell_predictor.ipynb. The model weights are in model_weights.pth. You can run the notebook to see the model training and evaluation process.

The notebooks prefixed with experiment_XX were experiments used to improve this process.

If you train and save a model with a different architecture to the original, you'll need to update the duplicate model definition in models.py so that the weights can be loaded to run the game.

Acknowledgements

The original inspiration for this can be traced back to Ollin Boer Bohan's blog post here: https://madebyoll.in/posts/game_emulation_via_dnn/

The core of the game was adapted from Matheus Gomes's pygame-life: https://github.com/matheusgomes28/pygame-life