Network Sensitivity Analysis

Abstract

Incremental as well as large change sensitivity computation is frequently required for design optimization and computer aided design. For probabilistic analysis need also exists for multi variable sensitivity computation. In general incremental sensitivity computation can be handled using sensitivity network, adjoint network and algebraic matrix techniques. With sensitivity and adjoint network approaches, the algebraic differential operator notation presents an integrated approach for derivation of results in frequency and time domain. The main difference in the two approaches consists in solution of all or one network variable with respect to one or all m parameters undergoing change. Further difference exists when considering higher order sensitivity co-efficients since use of sensitivity network approach for nth order co-efficients necessitates solution of all (n—1) order sensitivity networks but adjoint network technique entails solution of (m+2) networks of which one is original and (m+1) are adjoint networks with unit drivers. The adjoint networks results also facilitate extension to a variety of applications. The algebraic matrix approach as applicable to linear time invariant networks in frequency domain is based on sensitivity network concept but results in ex plicit sensitivity vectors corresponding to hybrid formulation. The latter formulation is also used to derive explicit expression for large change sensitivity in terms of a modified system matrix of order typically equal to number of varying parameters. Since the number of parameters undergoing change is generally small as compared to total number of components in a large system, the method results in computational saving.