Hierarchical Bayesian inverse reinforcement learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Inverse reinforcement learning (IRL) aims to infer reward functions from expert demonstrations.
• Challenges include the ill-posed nature of reward function inference and noisy, sub-optimal expert data.

Purpose of the Study:

• To develop a robust IRL framework capable of handling sub-optimal expert behavior.
• To improve the accuracy of reward function inference in complex scenarios.

Main Methods:

• A novel hierarchical Bayesian framework is proposed for IRL.
• This framework generalizes existing IRL algorithms and explicitly models expert sub-optimality.

Main Results:

• Experiments on synthetic data demonstrate the framework's effectiveness and robustness to noisy expert data.
• Application to real-world taxi GPS traces shows high accuracy in predicting driving behavior.

Conclusions:

• The proposed hierarchical Bayesian IRL framework offers a significant advancement in reward inference.
• It provides a powerful tool for understanding and predicting expert behavior, even with imperfect data.