An On-line Search for Optimum Operating Conditions of Noisy Systems

Optimum process operating conditions are often achieved by on-line computer programming. A general method has been developed which is capable of optimizing processes whose responses are corrupted by random error. The method is particularly useful for application to processes whose response functions contain sharp ridges. Application to ridge and non-ridge systems gives the same results in each case showing that the rate of convergence of the search is totally unaffected by the presence of a ridge. An accelarated gradient method is used which obtains its derivatives from a quadratic interpolation. The inability of the gradient method to move along a ridge is overcome by projecting a straight line through points on the ridge to obtain new starting points for a gradient search. Points on the ridge are located by finding the stationary point of the polynomial which is constructed by regression through points found by the gradient search. The method is applied to several test functions known to the literature—a fast convergence rate is obtained for a wide range of noise levels.