Reward function corresponding to human driving behavior

Determine a reward function that corresponds to human driving behavior for use in multi-agent reinforcement learning for autonomous driving, clarifying the target objective that accurately captures how people drive in interactive traffic scenarios.

Background

In the discussion of multi-agent reinforcement learning for driving, the paper notes that one approach to achieving human-compatible behavior is to encode human conventions via reward design.

However, the authors explicitly state that the correct reward function corresponding to human driving is not known, and they also caution that reward shaping can induce undesired behaviors.

This highlights a foundational open question about the specification of the optimization objective for learning agents that behave in a human-like manner in traffic.