Effects of Learning State Dependence of Reward Weights on Trust and Team Performance in a Human-Robot Sequential Decision-Making Task

In this paper, we evaluate two interaction strategies for a robot in a sequential decision-making task: one which uses a state-dependent reward function and the other that uses a state-independent (constant) reward function. Towards this, we present a study done on Amazon Mechanical Turk to learn the state-dependent reward function. Using this reward function, we compare the two strategies in simulation, where we also set the risk levels actively to induce a difference between the two strategies. Our results indicate that the interaction strategy using the state-dependent reward function results in better trust and team performance compared to that using the state-independent reward function, especially when more of the state space is explored. Thus, there may be merit in learning a more finegrained reward function for a robot interacting with a human. The results of this study provide a starting point for a future human-subjects study.