A distributed agent-based approach for simulation-based optimization

Structural design and optimization in engineering are increasingly addressing non‐standard optimization problems (NSPs). These problems are characterized by a complex topology of the optimization space with respect to nonlinearity, multimodality, discontinuity, etc. By that, NSP can only be solved by means of computer simulations. In addition, the corresponding numerical approaches applied often tend to be noisy. Typical examples for NSP occur in robust optimization, where the solution has to be robust with respect to implementation errors, production tolerances or uncertain environmental conditions. However, a generally applicable strategy for solving such problem categories always equally efficiently is not yet available. To improve the situation, a distributed agent‐based optimization approach for solving NSPs is introduced in this paper. The elaborated approach consists of a network of cooperating but also competing strategy agents that wrap various strategies, especially optimization methods (e.g. SQP, DE, ES, PSO, etc.) using different search characteristics. In particular, the strategy agents contain an expert system modeling their specific behavior in an optimization environment by means of rules and facts on a highly abstract level. Further, different common interaction patterns have been defined to describe the structure of a strategy network and its interactions. For managing the complexity of NSPs using multi‐agent systems (MASs) efficiently, a simulation and experimentation platform has been developed. Serving as a computational steering tool, it applies MAS technology and accesses a network of various optimization strategies. As a consequence, an elegant interactive steering, a customized modeling and a powerful visualization of structural optimization processes are established. To demonstrate the far reaching applicability of the proposed approach, numerical examples are discussed, including nonlinear function and robust optimization problems. The results of the numerical experiments illustrate the potential of the agent‐based strategy network approach for collaborative solving, where observed synergy effects lead to an effective and efficient solution finding.