Implementation of "Unsupervised Task Clustering for Multi-Task Reinforcement Learning", paper available here

We use sacred to configure and monitor our experiments.

The implementation is split into three main files:

• train_em.py for EM, SP and PPT
• train_multi_head.py for the Multi-Head experiments
• train_atari_em.py for all Atari experiments

The implementation of the Atari experiments is closely based on the Dopamine framework.

Usage

To use our code, please install dependencies

pip install -r requirements.txt

and run our experiments with

python train_em.py

To chose a different set of tasks, use the sacred CLI logic, for example to replicate our experiments on the pendulum task set use:

python train_em.py with environment=pendulum

Logging is done via the sacred logging interface, which stores the results in results/sacred/$EXPERIMENT_ID/. They can then be processed for example with incense.

Potential Issues

If the source code is not running as expected, please try to install some system packages that gym might require for the atari games.

apt-get install -y libglu1-mesa-dev libgl1-mesa-dev libosmesa6-dev xvfb ffmpeg curl patchelf libglfw3 libglfw3-dev cmake zlib1g zlib1g-dev swig