Apex is a small, modular library that contains some implementations of continuous reinforcement learning algorithms. Fully compatible with OpenAI gym.
Running experiments
Basics
Any algorithm can be run from the apex.py entry point.
To run PPO on a cassie environment,
python apex.py ppo --env_name Cassie-v0 --num_procs 12 --run_name experiment01
To run TD3 on the gym environment Walker-v2,
python apex.py td3_async --env_name Walker-v2 --num_procs 12 --run_name experiment02
Logging details / Monitoring live training progress
Tensorboard logging is enabled by default for all algorithms. The logger expects that you supply an argument named logdir
, containing the root directory you want to store your logfiles in, and an argument named seed
, which is used to seed the pseudorandom number generators.
A basic command line script illustrating this is:
python apex.py ars --logdir logs/ars --seed 1337
The resulting directory tree would look something like this:
trained_models/ # directory with all of the saved models and tensorboard logs
└── ars # algorithm name
└── Cassie-v0 # environment name
└── 8b8b12-seed1 # unique run name created with hash of hyperparameters
├── actor.pt # actor network for algo
├── critic.pt # critic network for algo
├── events.out.tfevents # tensorboard binary file
├── experiment.info # readable hyperparameters for this run
└── experiment.pkl # loadable pickle of hyperparameters
Using tensorboard makes it easy to compare experiments and resume training later on.
To see live training progress
Run $ tensorboard --logdir logs/
then navigate to http://localhost:6006/
in your browser
Cassie Environments:
Cassie-v0
: basic unified environment for walking/running policiesCassieTraj-v0
: unified environment with reference trajectoriesCassiePlayground-v0
: environment for executing autonomous missionsCassieStanding-v0
: environment for training standing policies
Algorithms:
Currently implemented:
- Parallelism with Ray
- GAE/TD(lambda) estimators
- PPO, VPG with ratio objective and with log likelihood objective
- TD3 with Parameter Noise Exploration
- DDPG
- RDPG
- ARS
- Entropy based exploration bonus
- advantage centering (observation normalization WIP)
To be implemented long term:
- SAC
- GPO
- NAF
- SVG
- I2A
- PGPE
- Value Distribution
- Oracle methods (e.g. GPS)
- CUDA support (should be trivial but I don't have a GPU to test on currently)
Maybe implemented in future:
Acknowledgements
Thanks to @ikostrikov's whose great implementations were used for debugging. Also thanks to @rll for rllab, which inspired a lot of the high level interface and logging for this library, and to @OpenAI for the original PPO tensorflow implementation. Thanks to @sfujim for the clean implementations of TD3 and DDPG in PyTorch. Thanks @modestyachts for the easy to understand ARS implementation.