This is the repo for the paper "REValueD: Regularised Ensemble Value-Decomposition for Factorisable Markov Decision Processes", to appear at ICLR 2024. The paper can be found here.
algorithms.py contains DecQN as a baseline with REValueD incorporating the ensemble -- this repo contains on single threaded variant for easier accessibility for experimentation.
environment_utils.py includes a gym wrapper for the DM Control Suite with wrappers that discretise the continuous action space.
TODO:
- Add regulariser to REValueD.
- Add prioritised replay buffer.
- If sufficient interest, release version using parallel workers in Ray.