Project_RL

Implementation of the Deep Deterministic Policy Gradient (DDPG) and Maximum A Posteriori Policy Optimization (MPO) Reinforcement Learning Algorithms for continuous control on OpenAI gym environments.

Prerequisites

To use the Algorithms will require python3 (>=3.6.5).

• This should install all dependencies and the packages (cloning -> into a directory of choice):

  git clone https://github.com/theogruner/rl_pro_telu
  cd rl_pro_telu
  pip install -e .

• Or they can be installed manually

  Installation gym: https://github.com/openai/gym

  Installation quanser: https://git.ias.informatik.tu-darmstadt.de/quanser/clients/tree/master

  Additionally PyTorch, Numpy, Tensorflow, TensorboardX are required.

Usage Examples

Note

The algorithms are intended for continuous gym environments !

DDPG

• Usage with provided noises (Ornstein Uhlenbeck, Adaptive Parameter)

  If you want to use DDPG with our pre-implemented noises you can simply do this via the main_ddpg.py file.

  • Training on the Qube-v0 (can be replaced with any environment id e.g. Levitation-v0) environment with default hyperparameters and Ornstein-Uhlenbeck noise, saving the model as furuta_model.pt and the log in furuta_model (saves by default as ddpg_model.pt, and log in a automatic generated directory). Saving/logging could be disabled with --no-save and --no-log

    python3 path/to/main_ddpg.py --env Qube-v0 --save_path furuta_model.pt --log_name furuta_log

  • Loading a saved model and evaluating it in the Qube-v0 environment. Number of episodes and their length for testing can be adapted with --eval_episodes and --eval_ep_length

    python3 path/to/main_ddpg.py --env Qube-v0 --no-train --eval --load furuta_model.pt

  • Training and evaluating are executed sequentially, so if u want to evaluate a just trained model, this will work:

    python3 path/to/main_ddpg.py --env Qube-v0 --train -eval --save_path furuta_model.pt --log_name furuta_log 

  • Loading a model and setting the train flag will continue training on the loaded model

    python3 path/to/main_ddpg.py --env Qube-v0 --train -eval --save_path furuta_model.pt --log_name furuta_log --load furuta_model.pt

  • Adapting hyperparameters is quite intuitive:

    python3 path/to/main_ddpg.py --env Qube-v0 --gamma 0.5 --tau 0.1 --batch_size 1024

  • For more information, e.g. about adaptable hyperparameters, use:

    python3 path/to/main_ddpg.py --help

• Usage with self defined noise

  To use a self defined noise you will need to write a script by yourself.

  • Make sure the script is the same directory as the ddpg package (only if you didn't install them with PyPI).

  • The noise should extend the Noise class in noise.py (contain a reset and iteration function)

  • Following examples cover previous examples functionality

    import gym    
    import quanser_robots
    
    from ddpg import DDPG
    from ddpg import OrnsteinUhlenbeck
    
    # create environment and noise
    env = gym.make('Qube-v0')
    action_shape = env.action_space.shape[0] 
    noise = OrnsteinUhlenbeck(action_shape)
    
    # setup a DDPG model w.r.t. the environment and self defined noise
    model = DDPG(env, noise, save_path="furuta_model.pt", log_name="furuta_log")
    
    # load a model
    model.load_model("furuta_model.pt")
    #trains the model
    model.train()
    #evaluates the model and returns a meaned reward over all episodes
    mean_reward = model.eval(episodes=100, episode_length=500)
    print(mean_reward)
    
    # always close the environment when finished 
    env.cose()

  • Setting Hyperparameters would look something like this:

    model = DDPG(env, noise, gamma=0.5, tau=0.1, learning_rate=1e-2)

• Using a model as a controller

  If you want to use your a model as a simple controller just call the model passing an observation:

  ctrl = DDPG(env, noise)
  ctrl.load('furuta_model.pt')
  
  while not done:
     env.render()
     act = ctrl(obs)
     obs, rwd, done, info = env.step(act)
  
  # always close the environment when finished
  env.close()

MPO

Using MPO is analogous to ddpg

• Use main_mpo.py instead of main_ddpg.py

• For information on the parameters you can set: python3 main_mpo.py --help

• Due to no noise needed writing an own script is not necessary but for completeness a little code snippet as example:

  import gym
  import quanser_robots
  from mpo import MPO
  
  env = gym.make('Qube-v0')
  model = MPO(env, save_path="furuta_model.pt", log_name="furuta_log")
  # continues like DDPG example ...

Logging

By default logging is enabled and safes the logs in the runs/ directory. Name of the logs can be set by the log_name parameter (--log_name argument). Inspecting them works with:

tensorboard --logdir=*/PATH/TO/runs*

This starts a local server, which can be accessed in the browser. Connecting to the server should result in something like this:

Open Source Infos

Contributing

Please read CONTRIBUTING.md for details on our code of conduct, and the process for submitting pull requests to us.

Authors

• theogruner - theogruner
• DariusSchneider - DariusSchneider

See also the list of contributors who participated in this project.

License

This project is licensed under the GNU GPL3 License - see the LICENSE file for details