A comparison of optimization and search methods for multidisciplinary design

Abstract

Typically, papers on multidisciplinary design focus on application of a specific optimization or search technique to a given design problem. Little general guidance is provided for selecting an appropriate optimization or search method for multidisciplinary design. This paper discusses several optimization and search methods for multidisciplinary design. These methods are calculus-based optimization, optimization of response surfaces, expert systems, simulated annealing, genetic optimization and neural networks. The discussion of each method includes a brief description, a summary of advantages and limitations, and references for further reading. The intent is to provide insight into the behavior and applicability of each method. The best choice of optimization/search method, or combination of methods, depends on the features of the given multidisciplinary design problem. Multidisciplinary design problems can have many challenging features. These include heterogeneous mixes of analysis codes, discrete design parameter values, nondifferentiable functions, expensive functions, large numbers of design variables and many local minimizers. In addition, the best choice of optimization/search method is dependent on the organization of the data communication and execution paths of the various disciplines. The above issues are all considered in discussing the optimizationJsearch methods. In addition, the problem of obtaining derivatives from the individual discipline analysis codes is examined in the section on calculusbased methods. After the optimization/search melhods are described, a summary is given indicating the applicable methods for several types of problems. The possibility of combining methods is considered as well.