Risk management for self-adapting self-organizing emergent multi-agent systems performing dynamic task fulfillment

The goal of self-adapting self-organizing emergent multi-agent systems applied to problems with dynamically appearing tasks is to reduce operation and design costs. This is accomplished through the design of autonomous agents, which interact to produce behaviors required for flexible and scalable operation. However, when combined with agent autonomy, emergent behaviors are unpredictable resulting in a lack of trust for applications desiring efficiency such as logistics. An additional consultation agent, known as an efficiency improvement advisor (EIA), has been shown to increase efficiency through autonomy preserving advice provided as exception rule adaptations to agents. The problem addressed in this paper is that, in order for EIA-adapted systems to be deployed, the stakeholders must be assured that the risks of both autonomous and adapted behavior are properly assessed and managed. This paper presents a complete framework for a risk-aware EIA (RA-EIA) which uses reflection in order to manage the risks associated with autonomous agents and prospective adaptations. Monte Carlo simulation is used to reduce the frequency of emergent misbehavior appearing during regular operation. Meanwhile, an exploratory testing method, termed evolutionary learning of event sequences, is used to deal with the possibility of severe emergent misbehavior as the result of an malicious adversary or a series of unfortunate events. The experimental evaluations and accompanying descriptive example, for the application area of logistics via pickup and delivery problems, demonstrate that the risk-aware adaptations provided from consultation with the RA-EIA agent allow the client system to be trusted for long-term independent and reliable operational efficiency.