This is the project repository of my team in the "Advanced Deep Learning for Robotics" course at TUM. Our project's topic is "Exploring Munchausen Reinforcement Learning" based on this paper.
For a detailed discussion, see the report and the final presentation.
- Create a virtual environment.
- Run
pip3 install -r requirements.txt
This repository is structured as follows:
-
The directories
M-DQN
andM-SAC
contain the implementations of the RL agents DQN and SAC extended with the Munchausen term, respectively. -
The directories
rl-baselines3-zoo
contains a copy of this repository, where we included the implementations of M-DQN so that we can easily train and test the M-DQN agent on benchmark environments and also compare it to other classical agents. To do so, just follow the steps described in the original repository and insertM-DQN
as the agent argument. -
The directory
particles-env
contains a modified version of this repository. The modified version contains code for a particles environment, where an agent wants to reach a goal, while avoiding obstacles. Besides, M-SAC agent is implemented and included in the code, so that it can be trained and compared to the classical SAC agent. -
The directory
action-gap
contains implementation of callbacks for experiment manager of rl-baselines3-zoo which logs action-gap for tensorboard.