| 摘要 |
A method of efficient computation of gradients of a merit function of a system includes the steps of: specifying at least one parameter for which the gradients with respect to the at least one parameter are desired; specifying the merit function of interest in terms of observable measurements of the system; either solving or simulating the system to determine values of the measurements; expressing the gradients of the merit function as the gradient of a weighted sum of measurements; forming an appropriately configured adjoint system; and either solving or simulating the adjoint system to simultaneously determine the gradients of the merit function with respect to the at least one parameter by employing a single adjoint analysis. Preferably, the system may be modeled by a set of equations comprising at least one of the following: a nonlinear set of equations, a linear set of equations, a set of linear partial differential equations, a set of nonlinear partial differential equations, a set of linear differential algebraic equations or a set of nonlinear differential algebraic equations. Further, the system of interest may be a network and, preferably, may be an electrical circuit. Still further, elements of the adjoint network and excitations of the adjoint network are determined in order to obtain the gradients of the merit function by employing a single adjoint analysis. It is to be appreciated that, in a preferred embodiment, the gradients of merit function are computed for the purpose of optimization and the merit function may be either a Lagrangian merit function or an augmented Lagrangian merit function.
|
