Offline Preference-Based Reward Learning

Daniel Shin, Anca D. Dragan, Daniel S. Brown

Overview

Codebase for Benchmarks and Algorithms for Offline Preference-Based Reward Learning.

Dependencies

To install relevant create dependencies, create conda environment with conda env create -f environment.yml Then install d4rl in the d4rl sub-directory.

Instructions

We provide code in three sub-directories: d4rl containing slightly modified version of the original d4rl dataset, reward_learning containing code for preference-based reward learning, and policy_learning for training offline reinforcement learning policies.

See corresponding READMEs in each folder for instructions; scripts should be run from the respective directories. It may be necessary to add the respective directories to your PYTHONPATH.