reinforcement-learning

Introduction

This repository houses a series of files meant to introduce the topic of reinforcement learning and apply it to a continuous state space to control a pendulum in OpenAI's Gymnasium. See the project demonstration here and learn more about my findings here.

File name	File type	Description
pendulum_deep.py	Main file	Main loop; constructs agent object, observes and manipulates environment
deep_agent.py	Agent class	Constructs network object, contains functions for choosing actions, storing memories, and learning
deep_network.py	Network class	Builds Torch neural network and contains functions for saving network weights for future evaluation
replay_buffer.py	Memory buffer class	Handles memory control, contains functions for appending to and sampling from memory
utils.py	Plotting tools	Modularizes plotting learning rates

Setup

Python 3.10 is used, but Python 3.8 also appears to work as intended.

Install libraries:

pip install torch matplotlib numpy gymnasium gymnasium[classic-control]

Navigate to project/ and run:

python3 pendulum_deep.py

Any changes to hyperparameters or the number of games to be run can be edited directly in pendulum_deep.py. By default, as the average score improves, the program saves both networks to the models/ folder, and upon completion, plots the training data in plots/.