Game of life as a neural network

A passing thought came to my mind - the rules of Game of life are essentially a 3x3 convolution filter.

That means I can train a really simple neural network to advance a game of life board by one step. Furthermore - since convolution is the only thing going on, I can train the model on tiny boards and the network will work on boards of any size - and on the GPU!

The only thing that was not easy is making the network run on a toroid board - meaning that the board wraps around itself. you can look at the call method in network.py to see how I did that

Usage

I included a requirements.txt file, it has everything that's needed and some things that are not needed, if you install it you'll have everything you need:

pip install -r requirements.txt

To see a glider gun running on the model with pre-trained weights, just run main.py:

python main.py

To train the model yourself use the -t or --train option:

python main.py -t

You can save the weights using -s or --save:

python main.py -t -s

If you want to save or load the weights to different file use the -w or --model-weights option.

This is how you save weights:

python main.py -t -s -w my_weights.ckpt

and this is how you run the model with your saved weights:

python main.py -w my_weights.ckpt

In case you don't have a GPU the model will take a long time to train. I have created a larger version of the model which is easier to train, you can train it using this command:

python main.py -t -p cpu-training

The model

The model structure is very simple - it's a 3x3 convolutional layer followed by a 1x1 convolutional layer. It's implemented using the subclass API of TensorFlow in order to implement to wraparound behaviour.

Play with it!

The included model trains on a 25600 5x5 random boards, the model has 2 sigmoid activated filters and is trained for 300 epochs.

The cpu-training model trains on a 5024 5x5 random boards, the model has 32 relu activated filters and is trained for 30 epochs

You can create your own training profile in the training_parameters.yaml to try and optimize the learning time and the running time.

You can then run your profile using this command:

python main.py -t -p PROFILE_NAME_HERE

If you get loss of less than 1e-3 game of life will run reliably.

You can also change the function in the end of main.py to do something different with the end results. It currently downloads a glider gun .cells file from conwaylife.com and displays it on a 150x120 board. But you can do anything you want with it.

Display

The model works incredibly fast - the display unfortunately doesn't - I used ChatGPT to write 2 displays for game of life, one in QT5 and one in pygame - both of them are a bit slow since they draw the entire board every time. Since writing an efficient game of life display is not very interesting IMO I decided to let it go. Both displays are included, the pygame one is the one that's used. Initially I used matplotlib to draw everything and the code for that is included in draw.py