Theory And Methodology Of Model-based Integrated AI Planning In Hybrid Systems

AI planning is an important area in Artificial Intelligence (AI) research, which aims to produce a sequence of actions, called a plan, according to initial world state and some given action modes with the assistance of computer automatically or semi-automatically to reach goal world state or achieve goal tasks. AI planning techniques have potential for application in intelligent logistics and transportation, intelligent manufacture, software engineering, industrial system control, military operation, etc.Hierarchical task network (HTN) planning is an important framework and approach proposed for planning in realistic domain. Directed by Intelligence Engineering theory and methodology, combining with application in vehicle application and power plant domain, this dissertation explores some engineering-oriented issues from theoretical and methodological aspects based on HTN planning framework.Firstly, the concept of model-based HTN planning is proposed. A framework called Structural Object-Attribute-Relation (SOAR) model is then proposed for HTN planning in object-oriented, component-based hybrid environment that can make the system model be well-organized and expressive for symbolic and numeric hybrid problem solving. Based on this framework, some model-based reasoning algorithms are presented. Experiments in vehicle application domains show that this framework not only can integrate multiple types of knowledge but also improve planning efficiency significantly.Secondly, frameworks for integrating HTN planning with hybrid systems are explored combining with simulation technique. An algorithm is proposed that integrate ordered task decomposition (OTD) algorithm with explicit simulation algorithm. Based on this framework, Failure Treatment Plan (FTP) generation problem are defined, and then a condensing system FTP generation system in power plant operation is implemented to illustrate the feasibility of the framework.Plan optimization and evaluation is an important function of planners. An efficient multiple plans generation algorithm called segmented back tracking is proposed that can fast find a number of feasible plans during HTN planning process, and can decouple plan generation and plan evaluation process to provide flexible support to plan optimization and evaluation. Since most modern HTN planning modeling languages are theoretical research oriented and based on logical and/or functional expressions, the model is not clear to understand, easy to modify, extend and manage for engineering oriented research. A modeling language, called Model-based Integrated Planning Language (MIPL), is proposed which features object-oriented, component-based and hierarchical modeling. It also supports symbolic & numeric hybrid reasoning, computing and HTN planning. Most benchmarks in the research project are encoded in this language.To promote knowledge engineering using in AI planning, an ontology for HTN planning, SOAR-HTN, is designed based on SOAR framework using Protege system and a graphical knowledge environment is implemented mainly based on the proposed ontology. Practice shows that, in some conditions, this approach features better model consistency, easy to interact and cooperate among planners.Finally, conclusions of all research efforts in this dissertation are made and further research suggestions are given.