My implementation of PPO based on OpenAI's baseline implementation.

• Gym wrappers

  • GymWrap: Ordinary wrapper for continuous environments like RoboSchool
  • StackObs: Used for partially observable environments in order to estimate velocity acceleration etc..
  • ResetEarly: Can be used combined with GRU or LSTM models in partially observable environments, however this training process is very slow.

• Models

  • MLP: Standard 2 hidden layers FC neural network, where the actor and critic network are independent.
  • CNN: Shared CNN-block followed by separate output layers for the actor and critic.
  • LSTM and GRU: Recurrent neural network models that can model partially observable environments without the help of stacked frames. (Note that these should be used when the episode steps < the minibatch size)

The models can be edited in ppo.py

main_atari.py will run Brakeout by default.
main_roboschool.py will run InvertedPendulum by default.

The environments and other hyperparameters can be changed in the tensorflow flags.