Implementation of a multi-processed Proximal Policy Optimization (PPO) with some "implementation tricks" from the article What Matters In On-Policy Reinforcement Learning?
It achieves a score of around 300 (considered solved) on the BipedalWalker-v2 environment from OpenAI:
Added the following features:
- Multi-processed workers for experience collection (resulting in 2-3x faster wall clock time).
- Model loading to allow warmstarting or continuing from a checkpoint.
- Proper seeding resulting in fully reproducible runs.
- Tensorboard logging.