Intro to Reinforcement Learning

Development of DQN using TensorFlow NNs

As preparation for my thesis, RL for a reservoir computing, I completed this project to solidify my understanding of deep reinforcement learning. This project was done directly after reading and studying corresponding chapters of “AI: A Modern Approach”.

Since DQN is the first algorithm I am to implement in my thesis, I focused on a standard implementation of the algorithm using TensorFlow for the neural network. Based on the paper, I implemented experience replay buffers, epsilon-greedy action selection and batch gradient descent. Additionally, I added reward shaping, intermediate model saving, and Weights & Biases integration.

Results

I tested my model on OpenAI Gymnasium’s Cart Pole and Mountain Car. I successfully played each, achieving strong learning performance with my algorithm.

Cart Pole

Mountain Car