Roadmap

• Rewrite core
• Distributed RL
  • Some References:
    • Ray Rllib: https://github.com/ray-project/ray/blob/master/doc/source/rllib.rst
      • papers: Ray RLlib: http://proceedings.mlr.press/v80/liang18b.html Ray: https://arxiv.org/abs/1712.05889
      • talks: Good Intro to distributed RL (first ~25 mins) then discussion of Ray RLlib (next ~15 mins, the rest of the talk is on imitation learning), by two of the RLlib authors: https://www.youtube.com/watch?v=Y6feXBY6_XQ&t=0s&list=PLkFD6_40KJIxJMR-j5A1mkxK26gh_qg37&index=6
      • blogs: https://rise.cs.berkeley.edu/blog/scaling-multi-agent-rl-with-rllib/
    • Ape-x: https://openreview.net/forum?id=H1Dy---0Z
      • uber single machine implementation: https://github.com/uber-research/ape-x
    • R2D2: Extension of Ape-x with recurrent networks: https://openreview.net/forum?id=r1lyTjAqYX
    • IMPALA: https://icml.cc/Conferences/2018/Schedule?showEvent=3093
    • Uber Faster Neuro-evolution: https://eng.uber.com/accelerated-neuroevolution/
    • DOPAMINE: A RESEARCH FRAMEWORK FOR DEEP REINFORCEMENT LEARNING. https://arxiv.org/pdf/1812.06110.pdf

      Conclusions. From the above taxonomy we conclude that there is a natural trade-off between simplicity and complexity in a deep RL research framework. The resulting choices empower some research objectives, possibly at the detriment of others.

• HyperParameter Optimisation
• Support for easily deploying experiments
• More environments
  • Bullet
  • Box2D