A two-stage algorithm is introduced to find the optimal (in the least-squares sense), linear approximate model of a non-linear steady-state composite system having a structure described by an interconnection matrix. First, an optimum linear model of a complex system as a whole is computed and then optimal parameters for the component submodels of the complex model maintaining the structure of the system, are derived in a decentralized manner. The particular form of the algorithm for series system is discussed separately. The practical utility of the approach is demonstrated by means of some examples.