This repository documents the work completed by Matthew Stewart, Claire Stolz, and Shane Ong for the final practical of CS181: Machine Learning, involving the application of reinforcement learning to the flappy bird-style game 'Swingy Monkey'.

In this documentation we outline the methods used to attain the highest possible score on the game by using Q-learning and its deep learning-based extensions, commonly referred to as DQN (Deep Q-Learning) and DDQN (Double Deep Q-Learning).

To install Swingy Monkey, follow the instructions below. For tips on installing pygame, see the pygame documentation.

pip install pygame

To run the individual models, first make sure you have pygame installed. Then use the commands for each of the individual models:

python Q-learning_model.py 
python DQN.py
python DDQN.py

The score for the three models developed in this practical are outlined in the table below.

Hyperparameter optimization was performed in order to optimize for the ε-greedy policy, the learning rate α, as well as the discount rate γ, and other factors such as the neural architecture and hyperparameters of the deep neural network in the deep learning models.