Multiagent robustness: autonomy vs. organisation

This thesis reports on work conducted under the Schwerpunktprogramm Sozionik,1 a basic research project funded by the Deutsche Forschungsgemeinschaft to transfer knowledge between sociology and distributed artificial intelligence (DAI). The motivation of this work is to use sociological notions and theories of ’organisation’ to blueprint more robust multiagent systems. Based on an analysis of the DAI literature, we give a precise and empirically verifiable definition of robustness, which we call τ-robustness. This notion consists of (i) the definition of a performance measure, (ii) the definition of a perturbation against which the performance measure is evaluated, and (iii) a threshold τ , which marks the maximally allowed deviation of the performance measure to call the tested system τ-robust. The main theoretic contribution of this thesis is a framework of design parameters for robust multiagent organisation based on a sociological notion of organisation. This framework, called the Framework for selfOrganisation and Robustness in Multiagent systems (FORM), can be applied to multiagent organisation using the concept of holonic multiagent systems, for which a new, extended definition is presented. By freely combining all possible values of the parameters in the framework, it allows to model more than 90 000 different forms of multiagent organisation. Furthermore, we discuss how the holonic design parameters relate to different dimensions of the autonomy of the participating agents. This discussion shows that, and how, the notion of self-organisation is connected to the notion of adjustable autonomy. As the notion of autonomy is central to the definition of an agent, this constitutes an important contribution to the general theory of multiagent systems. In order to make the large design space spanned by FORM concrete, we chose to model a subset of the possible organisational forms. This choice is based on a sociological analysis of organisations in today’s economy and makes a diverse use of the available values for the design parameters. The modelled forms of organisation are arranged on a spectrum of autonomy ranging from fully autonomous agents to more and more coupled agents, until we reach an organisational form where the boundaries between individual agents dissolve and only one single agent remains. This spectrum can be used to devise a mechanism for self-organisation in the sense that agents start to organise in a loosely coupled holon and increase the 1 Collaborative research group in Socionics.