An integrated approach to hierarchical planning, plan execution, and on-line learning in uncertain continuous domains.

摘要

Classical planning systems reason in abstract domains, ignoring many of the issues which complicate realistic applications. These issues must be addressed when implementing a practical planning system in a real-world application. This dissertation develops and analyzes an integrated approach to planning, plan-execution, and learning in a class of complex domains, motivated by the need to perform mission planning and intelligent control of an autonomous Hybrid Electric Vehicle (HEV).; The planning problem is formally defined in a function-based framework which models plan execution as a Markov process in a general state space. A domain taxonomy is proposed which classifies planning domains based on their complexity and the presence of uncertainty. Novel issues within each domain class are analyzed, emphasizing uncertain continuous (UC) domains. Planning in UC domains requires the ability to react to unexpected plan-execution scenarios caused by significant domain uncertainty. Planning systems from the literature are analyzed under the proposed domain taxonomy.; This analysis leads to specification of RIPPELS, the Real-time Integrated Planning, Plan-Execution, and Learning System. RIPPELS consists of 3 domain-independent subsystems, which perform planning, plan execution, and learning. These subsystems interact in a way that leads to robust planning and plan execution in UC domains. RIPPELS was developed to confirm the hypothesis that planning performance and plan quality can be improved by incorporating a learning capability when planning in UC domains.; RIPPELS performs on-line, hierarchical, nonlinear, least-commitment planning subject to resource constraints and an explicit utility function. An execution monitor detects unexpected situations, to which the system may respond by replanning. RIPPELS also monitors the accuracy of its domain-models. Inaccurate domain models represent learning opportunities. Such opportunities are identified on-line and addressed by the learning subsystem, to improve domain-model accuracy based on the observed state during plan execution. Learning is represented as function approximation.; Simulation experiments confirmed that the combination of replanning and learning capabilities leads to improved planning efficiency and plans of higher utility. These experiments used RIPPELS to perform energy management and route-planning for an autonomous HEV.