General-Sum Multi-Agent Continuous Inverse Optimal Control

Modeling possible future outcomes of robot-human interactions is of importance in the intelligent vehicle and mobile robotics domains. Knowing the reward function that explains the observed behavior of a human agent is advantageous for modeling the behavior with Markov Decision Processes (MDPs). However, learning the rewards that determine the observed actions from data is complicated by interactions. We present a novel inverse reinforcement learning (IRL) algorithm that can infer the reward function in multi-Agent interactive scenarios. In particular, the agents may act boundedly rational (i.e., sub-optimal), a characteristic that is typical for human decision making. Additionally, every agent optimizes its own reward function which makes it possible to address non-cooperative setups. In contrast to other methods, the algorithm does not rely on reinforcement learning during inference of the parameters of the reward function. We demonstrate that our proposed method accurately infers the ground truth reward function in two-Agent interactive experiments.1