Computational Interaction Frames
Michael Rovatsos · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2004
This thesis introduces interaction frames as a novel approach for social reasoning and interaction management in multiagent systems. Interaction frames are a sociologically inspired concept that can be used to represent categories of interaction patterns. Agents operating and interacting in multiagent systems can employ these frames to record their interaction experience and to make strategic communication decisions based on this experience. This kind of socially intelligent framing that combines decision-theoretic rationality with empirical methods for learning the regularities of communication processes in a multiagent system is particularly well-suited for open agent societies. In these societies, the adherence to an a priori semantics of messages and communication protocols cannot be taken for granted. In the absence of absolute certainty about the ways others will behave in future interactions, the frame-based approach relies on observation and adaptation of one's own behaviour to the observed patterns of interaction to reduce uncertainty. This constitutes a significant improvement over the use of pre-specified communication protocols and conversation policies in a hard-wired fashion which can be too limiting, at least if we take agent autonomy seriously. We present an abstract social reasoning architecture called InFFrA that is based on the concepts of interaction frames and framing and that can be used as a meta-model for concrete agent designs. This abstract architecture is supplemented by the formal model of a concrete, ready-to-implement instance of the meta-model that complies with InFFrA requirements. For this formal model, we also establish a formal semantics based on a more general model of empirical semantics for agent communication. Furthermore, we define decision-making procedures for this formal version of InFFrA and develop learning algorithms that borrow from the theory of hierarchical reinforcement learning. An implementation of the formal social reasoning architecture is used to evaluate the performance of frame-based agents in a realistic application scenario taken from the domain of agent-based Web linkage. The experimental results prove the effectiveness of our approach and show that interaction frames can be successfully used as a powerful tool for reasoning about interaction in open systems. Finally, the broad applicability of frames is illustrated by a discussion of various further applications.