Abstract: Metareasoning is the process by which an autonomous system optimizes, specifically monitors and controls, its own planning and execution processes in order to operate more effectively in its environment. As autonomous systems rapidly grow in sophistication and autonomy, the need for metareasoning has become critical for efficient and reliable operation in noisy, stochastic, unstructured domains for long periods of time. However, despite considerable progress in metareasoning as a whole over the last thirty years, techniques that improve planning rely on many assumptions that diminish their accuracy and practical utility while those that improve execution have not seen much attention yet. The goal of this dissertation is to develop metareasoning approaches that boost both the efficiency of planning and the reliability of execution in autonomous systems.
This dissertation therefore introduces a two-pronged framework that offers more effective metareasoning for planning and expands the scope of metareasoning to execution. We begin by offering two metareasoning approaches for efficient planning: (1) a method that estimates the optimal stopping point of an anytime planner by predicting its performance at runtime and (2) a method that estimates the optimal hyperparameters of an anytime planner by learning from experience with deep reinforcement learning. We then offer two metareasoning approaches for reliable execution: (3) a method that recovers from exceptional circumstances by interleaving a main task process with a set of exception handlers and (4) a method that maintains and restores safe operation by executing in parallel a main task process and a set of safety processes with a decision-theoretic algorithm for conflict resolution.
Advisor: Shlomo Zilberstein