Reinforcement Learning in OpenAI's CartPole environment

Final Project for the Reinforcement Learning Lecture

Arce y de la Borbolla, José
Sälinger, Andreas

This repository contains the code for training, testing and evaluating Reinforcement Learning agents in a custom version of OpenAI's CartPole environment with continuous action-space, overloaded reward-function and episode-termination conditions.

Currently, the following Reinforcement Learning algorithms have been developed:

• DQN with optional Double-Q and Replay Buffer
• REINFORCE with Gaussian, Beta and MLP policies
• TD3 with MLP Actor and Critics

Directories:

• agents : contains the executable scripts for setting-up, training and testing agents.
• lib : classes, objects and methods, including the main algorithms
• scripts: helper scripts for testing and evaluating the agents
• save: saved agent models, statistics, output logs, plots and TensorBoard graphs.
• slides: Presentation slides with the results