This project will design, implement, and evaluate a system that integrates planning, execution, and learning in real time. The four main objectives to this research are: (1) Develop a usable execution monitor that will control plan execution in real time; (2) Develop the methods to plan directly from episodes in memory;(3) Automate the acquisition of cases from execution; and (4) Develop techniques for learning using a weak domain theory. The system's domain is meal planning and preparation. The system will plan and assist in the preparation of meals, then incorporate the results of these into its case memory. A central concern here is to design a system that is efficient enough to operate in real time. This will be attained in part by placing computational restrictions on time-critical processes, and in part by using a case-based approach to plan generation and repair that use episodic representations of previous executions. The detail in the resultant plans and repairs simplifies both execution monitoring and case acquisition. Learning is done at noncritical times, and is done without a strong domain theory by attributing expectation failures to failures in the assumptions used to generate the expectations.
