FOCAL: Efficient Fully-Offline Meta-Reinforcement Learning Via Distance Metric Learning and Behavior Regularization

We study the offline meta-reinforcement learning (OMRL) problem, a paradigm which enables reinforcement learning (RL) algorithms to quickly adapt to unseen tasks without any interactions with the environments, making RL truly practical in many real-world applications. This problem is still not fully understood, for which two major challenges need to be addressed. First, offline RL usually suffers from bootstrapping errors of out-of-distribution state-actions which leads to divergence of value functions. Second, meta-RL requires efficient and robust task inference learned jointly with control policy. In this work, we enforce behavior regularization on learned policy as a general approach to offline RL, combined with a deterministic context encoder for efficient task inference. We propose a novel negative-power distance metric on bounded context embedding space, whose gradients propagation is detached from the Bellman backup. We provide analysis and insight showing that some simple design choices can yield substantial improvements over recent approaches involving meta-RL and distance metric learning. To the best of our knowledge, our method is the first model-free and end-to-end OMRL algorithm, which is computationally efficient and demonstrated to outperform prior algorithms on several meta-RL benchmarks.

We have provided a latest codebase of FOCAL due to some reported issues, please check it out =).

Installation

First install MuJoCo. For task distributions in which the reward function varies (Cheetah, Ant), install MuJoCo150 or plus. Set LD_LIBRARY_PATH to point to both the MuJoCo binaries (/$HOME/.mujoco/mujoco200/bin) as well as the gpu drivers (something like /usr/lib/nvidia-390, you can find your version by running nvidia-smi).

For the remaining dependencies, create conda environment by

conda env create -f environment.yaml

For Walker environments, MuJoCo131 is required. Simply install it the same way as MuJoCo200. To swtch between different MuJoCo versions:

export MUJOCO_PY_MJPRO_PATH=~/.mujoco/mjpro${VERSION_NUM}
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mjpro${VERSION_NUM}/bin

The environments make use of the module rand_param_envs which is submoduled in this repository. Add the module to your python path, export PYTHONPATH=./rand_param_envs:$PYTHONPATH (Check out direnv for handy directory-dependent path managenement.)

This installation has been tested only on 64-bit CentOS 7.2. The whole pipeline consists of two stages: data generation and Offline RL experiments:

Data Generation

FOCAL requires fixed data (batch) for meta-training and meta-testing, which are generated by trained SAC behavior policies. Experiments at this stage are configured via train.yaml located in ./rlkit/torch/sac/pytorch_sac/config/.

Example of training policies and generating trajectories on multiple tasks:

python policy_train.py --gpu 0

Generate trajectories from pretrained models

python policy_train.py --eval

Generated data will be saved in ./data/

Offline RL Experiments

Experiments are configured via .json configuration files located in ./configs. Basic settings are defined and described in ./configs/default.py. To reproduce an experiment, run:

python launch_experiment.py ./configs/[EXP].json

By default the code run on GPU. To use CPU instead, set use_gpu=False in the corresponding config file.

Output files will be written to ./output/[ENV]/[EXP NAME] where the experiment name corresponds to the process starting time. The file progress.csv contains statistics logged over the course of training. data_epoch_[EPOCH].csv contains embedding vector statistics. We recommend viskit for visualizing learning curves: https://github.com/vitchyr/viskit. Network weights are also snapshotted during training.

To evaluate a learned policy after training has concluded, run sim_policy.py. This script will run a given policy across a set of evaluation tasks and optionally generate a video of these trajectories. Rendering is offline and the video is saved to the experiment folder.

Example of running experiment on walker_rand_params environment:

• download walker data and unzip the data to ./data/walker_rand_params (Download all normalized offline training data used in the FOCAL paper here)
• edit walker_rand_params.json to add dump_eval_paths=1 and data_dir=./data/walker_rand_params
• run

python launch_experiment.py ./configs/walker_rand_params.json

Reproducing Result in FOCAL Paper

We provide code for reproducing figure 2-9 and table 1 in generate_plot.py. Use output data to download the output files required for visualization and add them to ./output/ directory. To produce all figures at a time, run

python3 generate_plot.py

To produce each figure individually, run the function named by the corresponding figure number in main().

References

@inproceedings{li2021focal,
  title={{FOCAL}: Efficient Fully-Offline Meta-Reinforcement Learning via Distance Metric Learning and Behavior Regularization},
  author={Lanqing Li and Rui Yang and Dijun Luo},
  booktitle={International Conference on Learning Representations},
  year={2021},
  url={https://openreview.net/forum?id=8cpHIfgY4Dj}
}