MineRL 2021 Research track baseline agent uses Behavioural cloning (BC).
The agent can be run in three different ways:
- Colab notebook
- Standalone file
- Submittable repository (Under construction)
This is the same code as in the standalone file, but with extra documentation and fully running in Colab:
(the only difference is that the DATA_SAMPLES parameter is lower to fit the Colab RAM limits)
These are used to run the agent on your local machine and are located in the standalone directory.
MineRL requires JDK8 to be installed first.
After that, run:
pip install --upgrade minerl
pip install pytorch
pip install stable_baselines3
The agent can be run with:
python Behavioural_cloning.py
The BC agent comes with pretrained models. If you want to train it yourself, you will have to uncomment the train() line at the end of the file.
The following link will lead you to submittable version of the baselines:
Baseline results over several hundreds of episodes can be seen below:
Results above are achieved using the provided trained models. Colab baselines retrain every time and use a lower DATA_SAMPLES parameter (400,000 instead of 1,000,000) to fit the available RAM. Here are the results from three separate fully retrained models using those parameters:
Thank you to Anssi Kanervisto for developing most of the Behavioural cloning baseline.