Graphical models: Inference, decision and acquisition

点击翻译按钮获取中文摘要

My research concerns intelligent systems, known as agents, that function in uncertain or constraint-based environments, either individually or cooperatively. It focuses on graphical models for knowledge representation and covers issues on knowledge acquisition, inference, and decision making. My previous research has established several classes of multiagent graphical models, known as MSBNs for probabilistic reasoning, DMSBNs for forecasting, and CDNs for design decision making. Agent communication in MSBNs and DMSBNs is initiated by one agent (the root), and its election incurs overhead. To improve flexibility, feasibility of an unrooted regime will be investigated. Multiagent planning will also be studied focusing on online planning, rather than commonly pursued offline policy making, to gain efficiency. My previous research on multiagent constraint graphical models, known as MSCNs, confirms that lessons learned from multiagent probabilistic graphical models can be usefully extended into distributed constraint satisfaction (DCSP) and optimization (DCOP). The proposed research will explore the structure embedded in lower level runtime representation of MSCNs for more efficient constraint reasoning. Motivated by certain unique, desirable computational properties of CDNs and MSCNs (relative to existing frameworks for DCOP), a multiagent graphical model for DCOP will be developed by generalizing CDNs and MSCNs. Previous research developed causal models, NIN-AND trees, for efficient acquisition of conditional probability tables (CPTs) in constructing Bayesian networks. These models extend the expressive power of existing models from reinforcing interaction to undermining and mixture of the two. The proposed research will investigate approximating an arbitrary CPT as an NIN-AND tree, algorithms to acquire these models by data mining, and direct incorporation of NIN-AND trees into inference to improve efficiency.