Important: Deprecated version. Update to latest arxiv version coming soon. Please use the resources in our main repository in the meanwhile.
RUDDER efficiently learns optimal policies in finite Markov decision processes with delayed rewards. With the following links you can find:
- Main repository: https://github.com/ml-jku/rudder
- Our RUDDER paper: https://arxiv.org/abs/1806.07857
- RUDDER blog: https://www.jku.at/index.php?id=16426
- Code for RUDDER demonstration on example-task in blog: https://github.com/ml-jku/rudder-demonstration-code
- A practical step-by-step guide to applying RUDDER in PyTorch: https://github.com/widmi/rudder-a-practical-tutorial
Baselines incl. RUDDER
Uses RUDDER on PPO to solve environments with delayed rewards. Our RUDDER paper is available at https://arxiv.org/abs/1806.07857.
See https://github.com/widmi/rudder-a-practical-tutorial for a quick tutorial on reward redistribution using a pytorch implementation and code-snippets that you can run on your CPU.
Our code is based on the OpenAI Baselines package, in which we included our implementation of RUDDER for PPO for ATARI games with delayed rewards. Additionally, the Tensorflow Layer Library (TeLL) package (v1.0.0) is required for RUDDER.
RUDDER for PPO2 and documentation for the configuration is located in the folder baselines/ppo2_rudder.
Videos are available at: