Opportunistic plan execution monitoring and control
When executing a plan, the traditional assumption of complete certainty is rarely valid; in real-world situations, actions may fail to complete, or may take much longer than initially anticipated. Because of this, layers of robustness are required that allow flexibility during execution. There are two aspects of an executive that need addressing to provide the ability to deal with uncertainties: monitoring and control. This thesis examines both of these aspects and shows how these may be used during execution. For monitoring, the use of Hidden Markov Models is proposed. Tracing the execution of a task through a model allows the executive to detect its current state, even when this is not directly observable from within the system. The trajectory of states through the model can reveal execution irregularities, which can be flagged and identified as failure. This work explores the use of learnt models, evaluating their performance in various contexts. For control, the use of opportunistic plans is examined. These can be used after failure has occurred, or when an action successfully finishes early. The definition of opportunistic plans is expanded to include extra structures that may be of use during the creation of these. Execution issues surrounding opportunistic plans are addressed, including how to decide when to execute an opportunity. The logical requirements for this are discussed, and several opportunity insertion strategies are developed. This is followed by a technique for deciding if resources should be conserved for later opportunities with the prospect of greater reward.