Interactive Multiagent Adaptation of Individual Classification Models for Decision Support

Abstract

An essential prerequisite for informed decision-making of intelligent agents is direct access to empirical knowledge for situation assessment. This contribution introduces an agent-oriented knowledge management framework for learning agents facing impediments in self-contained acquisition of classification models. The framework enables the emergence of dynamic knowledge networks among benevolent agents forming a community of practice in open multiagent systems. Agents in an advisee role are enabled to pinpoint learning impediments in terms of critical training cases and to engage in a goal-directed discourse with an advisor panel to overcome identified issues. The advisors provide arguments supporting and hence explaining those critical cases. Using such input as additional background knowledge, advisees can adapt their models in iterative relearning organized as a search through model space. An extensive empirical evaluation in two real-world domains validates the presented approach.