Synthesising Liberal Normative Systems

Abstract

Norms have been extensively studied to coordinate multi-agent sys-tems, and the literature has investigated two general approaches to norm synthesis: off-line (synthesising norms at design-time) and on-line (run-time synthesis). On-line synthesis is generally recog-nised to be appropriate for open systems, where aspects of the system remain unknown at design-time. In this paper we present LION, an algorithm aimed at synthesising liberal normative sys-tems. LION’s normative systems respect the agents ’ autonomy to the greatest possible extent, constraining their behaviour when only necessary to avoid undesirable system states. LION’s norm synthe-sis is also driven by the need to construct compact normative sys-tems. The key to the success of LION in this multi-objective synthe-sis process is that it learns about and exploits norm synergies. More precisely, LION can learn when norms are either substitutable or complementary. We show empirically that LION significantly out-performs the state of the art by synthesising normative systems that are more liberal while maintaining representation compactness.