Important: visit https://sites.google.com/view/meta-acl for videos !
1- Unzip and install codebase with Python >= 3.6, using Conda for example
cd meta-ACL/
conda create --name metaACL python=3.6
conda activate metaACL
pip install -e .
2 versions of the toy environment can be used.
python3 toy_run.py --seed 42 --exp_name test_toy_env --teacher ALP-GMM -rnd 10
python3 toy_run.py --seed 42 --exp_name test_full_toy_env -v2 --teacher ALP-GMM -rnd 10 -rsc
python3 vizu_walker_climber_environment.py
python3 run.py --seed 42 --exp_name test_parkour --teacher ALP-GMM -rnd 10 --nb_test_episodes 225 -rndls
To test AGAIN, one first has to train a batch of students with ALP-GMM (i.e. a classroom). The resulting runs' data (located in teachDRL/data//) then has to be processed to generate a classroom file that will be given to AGAIN (with the -cf argument).
Here is the process detailed for the regular toy environment:
python3 toy_run.py --seed 0 --exp_name test_toy_env --teacher ALP-GMM -rnd 10
python3 toy_run.py --seed 1 --exp_name test_toy_env --teacher ALP-GMM -rnd 10
python3 toy_run.py --seed 2 --exp_name test_toy_env --teacher ALP-GMM -rnd 10
...
Use toy_env_classroom_maker.ipynb to generate the classroom (i.e. toy_classroom.pkl) You also need to move your classroom data to a new folder:
mv teachDRL/data/<your-alpgmm-exp-name>/ teachDRL/data/elders_knowledge/
python3 toy_run.py --seed 42 --exp_name test_metaACL_toy_env --teacher AGAIN --expert_type R --use_alpgmm -rnd 2 --nb_cubes 20 -pt 2 -sR -cf toy_classroom
Aknowledgment: Our approach and our environments are implemented on top of two main codebases:
https://github.com/openai/spinningup
https://github.com/flowersteam/teachDeepRL
Many thanks to their respective authors.