Relevanzbasierte Informationsbeschaffung für die informierte Entscheidungsfindung intelligenter Agenten

Abstract

This dissertation introduces relevance-based information acquisition for intelligent software agents based on Howard s information value theory and decision networks. Active information acquisition is crucial in domains with partial observability in order to establish situation awareness of autonomous systems for deliberate decisions. The new semi-myopic approach addresses the complexity challenge of decision-theoretic relevance computation by reducing the set of variables to be evaluated in the first place. Links in a decision network encode stochastic dependencies of variables. Through utility dependency analysis using Pearl s d-separation criterion, the set of relevant variables can be efficiently reduced to a proven minimum without actually computing information value. In addition to an implementation with detailed runtime performance analysis, the applicability of the approach is shown in the domain of intelligent logistics control.