When coordinating their actions to accomplish a mission, the agents in a multi-agent system may use a collaboration algorithm to determine which agent performs which task. This paper describes a novel data-driven metareasoning approach that generates a metareasoning policy that the agents can use whenever they must collaborate to assign tasks. This metareasoning approach collects data about the performance of the algorithms at many decision points and uses this data to train a set of surrogate models that can estimate the expected performance of different algorithms. This yields a metareasoning policy that, based on the current state of the system, estimated the algorithms’ expected performance and chose the best one. For a ship protection scenario, computational results show that one version of the metareasoning policy performed as well as the best component algorithm but required less computational effort. The proposed data-driven metareasoning approach could be a promising tool for developing policies to control multi-agent autonomous systems.
