This repository primarily consists of
- An implementation of the original Taxi domain [Dietterich, JAIR2000] with additional environmental complexities. Built on David Abel's simple_rl framework.
- Methods of a) selecting demonstrations that effectively summarize the agent's policy (i.e. behavior) to a human, and b) requesting demonstrations of what the human believes an agent would do in specific environments.
Required packages include numpy, matplotlib, pypoman to perform computational geometry with polytopes (i.e. BEC regions, see below), and pygame if you want to visualize some MDPs.
The main file is augmented_taxi.py, which currently has functions for a) generating an agent in the Augmented Taxi MDP, b) obtaining a Bayesian IRL summary of the agent's policy, c) obtaining a BEC summary of the agent's policy, and d) obtaining test environments to query the human's understanding of the agent's policy.