A Kalman-tracking filter approach to nonlinear programming (Q750314)

The problem considered is that of finding the minimum of a nonlinear function f(\(\theta\)) when the gradient and Hessian matrix are unknown or cannot be easily computed. The function f(\(\theta\)) may also be observed in the presence of stochastic noise. This type of nonlinear programming problem is here associated with a tracking problem arising in aerospace applications. At this time, finding the minimum of the function f(\(\theta\)) is equivalent to tracking a point \(\theta^*\) (the minimum) in the parameter space, based on noisy measurements of its position (the function evaluations f(\(\theta\))). A new nonlinear programming algorithm motivated by this tracking problem is presented which accounts for both modeling and approximation errors. The algorithm uses a two-level Kalman filter. The low-level Kalman filter performs function approximations, while the upper-level filter accounts for the tracking of parameters. The numerical examples shown in the paper indicate that the algorithm is robust to noise and is comparable to Newton's method.