Intelligent Cooperative Control Architecture: A Framework for Performance Improvement Using Safe Learning

摘要

Planning for multi-agent systems such as task assignment for teams of limited-fuel unmanned aerial vehicles (UAVs) is challenging due to uncertainties in the assumed models and the very large size of the planning space. Researchers have developed fast cooperative planners based on simple models (e.g., linear and deterministic dynamics), yet inaccuracies in assumed models will impact the resulting performance. Learning techniques are capable of adapting the model and providing better policies asymptotically compared to cooperative planners, yet they often violate the safety conditions of the system due to their exploratory nature. Moreover they frequently require an impractically large number of interactions to perform well. This paper introduces the intelligent Cooperative Control Architecture (iCCA) as a framework for combining cooperative planners and reinforcement learning techniques. iCCA improves the policy of the cooperative planner, while reduces the risk and sample complexity of the learner. Empirical results in gridworld and task assignment for fuel-limited UAV domains with problem sizes up to 9 billion state-action pairs verify the advantage of iCCA over pure learning and planning strategies.