MetaGenRL: Improving Generalization in Meta Reinforcement Learning using Learned Objectives
This is the official research code for the paper Kirsch et al. 2019: "Improving Generalization in Meta Reinforcement Learning using Learned Objectives".
Install the following dependencies (in a virtualenv preferably)
pip3 install ray gym[all] tensorflow-gpu scipy numpy
We also make use of ray's native tensorflow ops. Please compile them by running
python3 -c 'import ray; from pyarrow import plasma as plasma; plasma.build_plasma_tensorflow_op()'
Adapt the configuration in
ray_experiments.py (or use the default configuration) and run
python3 ray_experiments.py train
By default, this requires a local machine with 4 GPUs to run 20 agents in parallel. Alternatively, skip this and download a pre-trained objective function as described below.
After running meta-training (or downloading a pre-trained objective function)
you can train a new agent from scratch on an environment of your choice.
Optionally configure your training in
ray_experiments.py, then run
python3 ray_experiments.py test --objective TRAINING_DIRECTORY
This only requires a single GPU on your machine.
Using a pre-trained objective function
Download a pre-trained objective function,
cd ~/ray_results/metagenrl curl URL_COMING_SOON|tar xvz
and proceed with meta testing as above.
In this case your
TRAINING_DIRECTORY will be
Many tf summaries are written during training and testing and can be visualized with tensorboard
tensorboard --logdir ~/ray_results/metagenrl