Uncertainty propagation in multi-agent systems for multidisciplinary optimization problems

Abstract : Because of uncertainties on models and variables, deterministic multidisciplinary optimization may achieve under-sizing (without design margins) or over-sizing (with arbitrary design margins). Thus, it is necessary to implement multidisciplinary optimization methods that take into account the uncertainties in order to design systems that are both robust and reliable. Probabilistic methods such as reliability-based design optimization (RBDO) or robust design methods, provide designers with powerful decision-making tools but may involve very time-consuming calculations. New optimization approaches have been developed to deal with such complex problems. Auto-adaptive Multi-Agent Systems (AMAS) is a new approach developed recently, allowing to take into account the various aspects of a multidisciplinary optimization problem (multi-level, computation burden etc.). This approach was suggested for solving complex deterministic optimization problem. Now, the question of the integration of uncertainties in this multi-agent based optimization arises. The aim of this paper is to propose a new methodology for integrating the treatment of uncertainties in an adaptive multi-agent system for sequential optimization. The developed method employs a single loop process in which cycles of deterministic optimization alternate with evaluations of the system reliability. For each cycle, the optimization and the reliability analysis are decoupled from each other. The reliability analysis is carried out at agent level and only after the resolution of the deterministic optimization, to verify the feasibility of the constraints under uncertainties. Following the probabilistic study, the constraints violated (with low reliability) are shifted to the area of feasibility by integrating adaptive safety coeficients whose calculations are based on the agent-level reliability information. The method developed is applied to a conceptual aircraft design problem.