| Despite numerous applications and research in knowledge-based model construction techniques, there are two major and essential aspects that have not been addressed in conjunction: the formal properties of construction procedures and the efficiency of constructed models. This thesis addresses these two issues by providing a flexible representation language, sound procedures and applications to an important research area of Artificial Intelligence: the area of decision-theoretic planning.;To arrive at that objective, we propose and investigate an extension of logic programming, conditional probabilistic logic programming, which can be used to describe probabilistic and logical relationships and which allows the construction of Bayesian Networks. We address the theoretical, procedural, and application aspects of the framework. We define the declarative semantics of the language. We present a query-answering procedure and show its formal properties.;Second, we use the proposed language to represent action and domain knowledge for problems of planning under uncertainty. We investigate a knowledge-based model construction approach to contingent probabilistic plan evaluation. To evaluate contingent probabilistic plans, we propose the concept of Bayesian Network-graphs which are compact representations of inter-related Bayesian Network-fragments. We provide a procedure to construct a tailored Bayesian Network-graph to evaluate a given contingent probabilistic plan and investigate its formal properties.;Finally, we present an implemented decision-theoretic planner which functions on a probabilistic knowledge base of action and domain models. We assume that the plan spaces are specified by the user using a simple programming language. Our planner searches through the plan space to identify the optimal plan in the case of a single utility function and the set of undominated plans in the case of multiple utility functions. |
