Mündliche Doktorprüfung von Herrn Dipl.-Inf. Peter Novák am Freitag, 11.09.2009, 9 Uhr s.t. in Seminarraum 210
Behavioural State Machines - Agent Programming and Engineering
Abstract
One of the original long-term aims of Artificial Intelligence is to build intelligent entities, i.e., such which are able to function and act in a manner similar to human beings. Intelligent agents are autonomous entities, which are also assumed to be proactive, reactive, as well as socially able. To enable development of such systems, in this thesis, I am focusing on methods supporting design and implementation of cognitive agents, a specific subtype of agent systems which internally construct a model of their environment, themselves or their peers and use it as a basis for decisions about their future actions.
Since different application domains require different knowledge representation approaches, one of the outstanding problems of the field is the integration of heterogeneous knowledge representation approaches into a single agent system. I propose a framework of Behavioural State Machines (BSM), drawing a strict distinction between the representational and behavioural layer of an agent system. While keeping the former as abstract and open as possible, its focus lies on enabling flexible specification of agent’s behaviours and the overall agent reasoning model, i.e., specification of its lifecycle. Thus, the BSM framework features a plug-in architecture and enables transparent integration of heterogeneous knowledge representation technologies into a single agent system.
Provided a suitable programming framework, the next problem is the pragmatics of its use. Formal specification and semantic characterization of subprograms is essential means for getting a grip on composition of complex programs from lower-level modules. I introduce Dynamic Computation Tree Logic DCTL*, a novel hybrid logic marrying branching time temporal logic CTL* with features inspired by Dynamic Logic. It allows semantic characterization of BSM subprograms and in turn enables construction of a sequel of formally specified design patterns implementing useful agent-oriented concepts, such as achievement and maintenance goals. The thesis culminates in a proposal of commitment-oriented programming, an abstract methodology for programming cognitive agents generalizing the approach taken by construction of the design patterns.
To demonstrate the feasibility and usefulness of the proposed approach to agentoriented programming I describe three case-studies implemented in Jazzyk, a concrete implemented programming language for the BSM framework. Finally, to show the robustness of the BSM framework, I describe its straightforward probabilistic extension and present a formal study of its relationship to an agent-oriented programming language GOAL. I show that BSM can be seen as an intermediary language into which programs of other agent programming languages can be compiled.