Committee: 

Dr. Gombolay, Advisor

Dr. Klein, Co-Advisor    

Dr. Chernova, Chair

Dr. Egerstedt

Abstract: The object of the proposed research is to develop a novel graph attention network-based framework to automatically learn scalable scheduling policies for resource optimization. We aim to tackle problems in the challenging stochastic environments, with two scenarios being considered. First, we consider scheduling with stochastic and dynamic task completion times in human-robot team coordination—we extend the multi-robot task scheduling problem by introducing human co-workers. Heterogeneous task completion time is considered for human and robot workers represented by different probabilistic distributions. Second, we consider scheduling with stochastic and dynamic task arrival and completion times in failure-predictive plane maintenance—duration of the plane repair or maintenance task is modeled by a probability distribution affected by the plane status. Furthermore, the policy needs to schedule under the uncertainty of plane failure before a repair task is issued that greatly influences the future repair cost. We propose to use imitation learning to learn from imperfect demonstrations and further improve the model performance through policy-based reinforcement learning. By parameterizing the learner with graph attention networks, our framework is computationally efficient and results in a scalable resource optimization scheduler.