In this repository, we provide the code for our tutorial on applied reinforcement learning. We utilize TensorFlow's TF-Agents library to build a neural network agent capable of playing the video game Doom from pixels. To train the agent, Proximal Policy Optimization (PPO) is used.
For more details, have a look at our article.
This repository contains the following main components:
- ppo_train_eval_doom_simple.py: A minimal example on how to train Doom with TF-Agents and PPO.
- ppo_train_eval_doom_extended.py: A full example on how to train Doom with TF-Agents and PPO. This also includes logging metrics of training performance with TensorBoard and saving checkpoints.
- doom/DoomEnvironment.py:
An implementation of TF-Agents'
PyEnvironmentmapping running a vizdoom instance and mapping actions and the observations for our agent. - basic.cfg: Configuration for the basic Doom scenario with adaptions to work with our setup.
Please refer to our instructions in the article.