Dependable Risk-Aware Efficiency Improvement for Self-Organizing Emergent Systems

An efficiency improvement advisor agent acts as a consultation service for a self-organizing multi-agent system that improves operational efficiency. It identifies recurrent tasks in past problems that allow the creation of so-called exception rules for individual agents to limit future inefficient behavior. There exists the danger that introduced rules could possibly infringe on the flexibility and therefore reliability of the system. In this paper, we present a dependable risk-aware efficiency improvement advisor that uses Monte Carlo simulation techniques in strategic analysis assessing the long-term potential and risks of prospective rules. Our experimental evaluation, for the domain of dynamic pickup and delivery problems, shows that the result is a minimal, yet effective, set of risk-averse exception rules. These rules can be provided to individual agents to reliably achieve an overall long-term improvement in efficiency while maintaining flexibility.