This package provides a framework of different derivative-free optimizers (powered by Tensorflow 2.0.0) which can be used in conjuction with an MPC (model predictive controller) and an analytical/ learned dynamics model to control an agent in a gym environment.

The package features other functionalities to aid in model-based reinforcement learning (RL) research such as:

• Parallel implementation of the different optimizers using Tensorflow 2.0
• Loading/ saving system dynamics model.
• Monitoring progress using tensorboard.
• Learning dynamics functions.
• Recording videos.
• A modular and flexible interface design to enable research on different trajectory evaluation methods, optimizers, cost functions, system dynamics network architectures or even training algorithms.

Optimizers references:

• CEM
• CMA-ES
• PI2
• PSO
• SPSA

pip install blackbox_mpc

git clone https://github.com/ossamaAhmed/blackbox_mpc.git
cd blackbox_mpc
pip install -e .

pip install tensorflow_gpu==2.0.0

The easiest way to get familiar with the framework is to run through the tutorials provided. An example is shown below:

from blackbox_mpc.policies.mpc_policy import \
    MPCPolicy
from blackbox_mpc.utils.pendulum import PendulumTrueModel, \
    pendulum_reward_function
import gym

env = gym.make("Pendulum-v0")
mpc_policy = MPCPolicy(reward_function=pendulum_reward_function,
                       env_action_space=env.action_space,
                       env_observation_space=env.observation_space,
                       true_model=True,
                       dynamics_function=PendulumTrueModel(),
                       optimizer_name='RandomSearch',
                       num_agents=1)

current_obs = env.reset()
for t in range(200):
    action_to_execute, expected_obs, expected_reward = mpc_policy.act(
        current_obs, t)
    current_obs, reward, _, info = env.step(action_to_execute)
    env.render()

An API specification and explanation of the code components can be found here.

blackbox_mpc is work done by Ossama Ahmed (ETH Zürich), Jonas Rothfuss (ETH Zürich) and Prof. Andreas Krause (ETH Zurich).

This package was developed at the Learning and Adaptive Systems Lab @ETH Zurich.

@misc{blackbox_mpc,
   author = {Ahmed, Ossama and Rothfuss, Jonas and Krause, Andreas},
   year = {2020},
   publisher = {GitHub},
   journal = {GitHub repository},
   howpublished = {\url{https://github.com/ossamaAhmed/blackbox_mpc}},
}

The code is licenced under the MIT license and free to use by anyone without any restrictions.

• Add bayesian neural networks (BNN) and graph neural networks (GNN) support.
• Add different trajectory evaluators to propagate uncertainities support.