Code for performing Hierarchical RL based on the following publications:

"Data-Efficient Hierarchical Reinforcement Learning" by Ofir Nachum, Shixiang (Shane) Gu, Honglak Lee, and Sergey Levine (https://arxiv.org/abs/1805.08296).

"Near-Optimal Representation Learning for Hierarchical Reinforcement Learning" by Ofir Nachum, Shixiang (Shane) Gu, Honglak Lee, and Sergey Levine (https://arxiv.org/abs/1810.01257).

Requirements:

• TensorFlow (see https://www.tensorflow.org for how to install/upgrade)
• Gin Config (see https://github.com/google/gin-config)
• Tensorflow Agents (see https://github.com/tensorflow/agents)
• OpenAI Gym (see https://gym.openai.com/docs, be sure to install MuJoCo as well)
• NumPy (see https://www.numpy.org/)

Quick Start:

Run a training job based on the original HIRO paper on Ant Maze:

python scripts/local_train.py test1 hiro_orig ant_maze base_uvf suite

Run a continuous evaluation job for that experiment:

python scripts/local_eval.py test1 hiro_orig ant_maze base_uvf suite

To run the same experiment with online representation learning (the "Near-Optimal" paper), change hiro_orig to hiro_repr. You can also run with hiro_xy to run the same experiment with HIRO on only the xy coordinates of the agent.

To run on other environments, change ant_maze to something else; e.g., ant_push_multi, ant_fall_multi, etc. See context/configs/* for other options.

Basic Code Guide:

The code for training resides in train.py. The code trains a lower-level policy (a UVF agent in the code) and a higher-level policy (a MetaAgent in the code) concurrently. The higher-level policy communicates goals to the lower-level policy. In the code, this is called a context. Not only does the lower-level policy act with respect to a context (a higher-level specified goal), but the higher-level policy also acts with respect to an environment-specified context (corresponding to the navigation target location associated with the task). Therefore, in context/configs/* you will find both specifications for task setup as well as goal configurations. Most remaining hyperparameters used for training/evaluation may be found in configs/*.

NOTE: Not all the code corresponding to the "Near-Optimal" paper is included. Namely, changes to low-level policy training proposed in the paper (discounting and auxiliary rewards) are not implemented here. Performance should not change significantly.

Maintained by Ofir Nachum (ofirnachum).