Counterpropagation neural networks in decomposition based optimal design

Abstract The present paper explores a decomposition based approach in the optimal design of large-scale structural systems. The use of formal optimization methods in such systems is complicated by the presence of a large number of design variables and constraints. Decomposition reduces the large scale system into a sequence of smaller, more tractable subsystems, each with a smaller set of design variables and constraints. The decomposed subsystems, however, are not totally decoupled and design changes in one subsystem may have a profound influence on changes in other subsystems. The present work examines the effectiveness of counterpropagation (CP) neural networks as a tool to account for this coupling. This capability derives from a pattern completion capacity in such networks. The proposed approach is implemented for a class of structural design problems where the decomposed subsystems exhibit hierarchy, i.e. there is a distinct chain of command in the nature of couplings between the subsystems. Numerical results compare well with solutions obtained through the use of a traditional optimization implementation with no problem decomposition.