Fostering resilient execution of multi-agent plans through self-organisation

Abstract

Traditional multi-agent planning addresses the coordination of multiple agents towards common goals, by producing an integrated plan of actions for each of those agents. For systems made of large numbers of cooperating agents, however, the execution and monitoring of a plan should enhance its high-level steps, possibly involving entire sub-teams, with a flexible and adaptable lower-level behaviour of the individual agents. In order to achieve such a goal, we need to integrate the behaviour dictated by a multi-agent plan with self-organizing, swarm-based approaches, capable of automatically adapting their behaviour based on the contingent situation, departing from the predetermined plan whenever needed. Moreover, in order to deal with multiple domains and unpredictable situations, the system should, as far as possible, exhibit such capabilities without hard-coding the agents behaviour and interactions. In this paper, we investigate the relationship between multi-agent planning and self-organisation through the combination of two representative approaches both enjoying declarativity. We consider a functional approach to self-organising systems development, called Aggregate Programming (AP), and propose to exploit collective adaptive behaviour to carry out plan revisions. We describe preliminary results in this direction on a case study of execution monitoring and repair of a Multi-Agent PDDL plan.