colurw / snake_ppo

Nokia's classic 'snake' game, written in NumPy and converted into a Gymnasium Environment() for use with gradient-based reinforcement learning algorithms

snake_PPO

A work in progess....

Proximal Policy Optimisation is a state-of-the-art reinforcement learning algorithm designed to allow gradient learning methods to be applied to situations where an 'intelligent' agent makes autonomous decisions in an environment. This permits a back-propogation of the environment's gradients through the network (effectively) after each decision.

This (theoretically) should speed up the learning process when compared to gradient-free methods, such as flight_sim_GNN, which suffer through a randomised exploration of the environment's reward manifold.

snake_cli.py

Is an ascii version of the classic Nokia mobile game, playable in the command line using keyboard controls, that captures screen and keypress datastreams. Intended for a supervised learning project, that was abandoned after realising that I didn't want to spend hours playing it to create the data. Hence...

snake_pyg_num.py

Deprecates the data capture, and transforms the output into a RGB image array, and uses PyGame to generate a more attractive display window.

snake_gym.py

Refactors the game into a Gymnasium Environment() to simplify integration with reinforcement learning algorithms.

PPO.py

Initialises a convolutional neural network that learns how to use the control inputs, based on its interpretation the RGB screen array.

To Do

• Customise CNN policy to suit smaller screen.
• Integrate with 'Weights And Biases' or Tensorboard to aid with determining proper hyperparameters
• Capture PyGame screens to show learning progress over many epochs.