An Implementation of PILCO: Probabilistic Inference for Learning Control on the Pendulum environment

We recommend using a conda virtual environment. Installation assumes that the user already has a MuJuCo licence available at: https://www.roboti.us/license.html

conda create -n pilco python==3.7
conda activate pilco

pip install tensorflow==2.2.0
pip install gpflow==2.0.0
pip install gast==0.3.3
pip install gym
pip3 install -U 'mujoco-py<2.1,>=2.0'
conda install matplotlib pandas pytest

The example may be run as follows:

python pendulum_example.py

where the default hyperparameter values have been configured to produce a plot resembling the one below. Random seeds are not fixed and so different plots will be obtained every time the script is run. There is considerable scope for hyperparameter tuning!

The following excellent open-source projects: https://github.com/nrontsis and https://github.com/aidanscannell/pilco-tensorflow proved to be highly valuable references.