Analogue computer study of the effect of input spectrum on a hill-climbing optimizer

One of the techniques used in the optimization of controlled plant is that of hill-climbing. This technique is most usefully employed on-line and, therefore, must operate with the normal random operating signals. The performance (as opposed to economic) criterion of the plant´s operation is usually chosen to be the mean square error. This paper investigates the variations in the shape of the curves of mean square error against the forward, and rate feedback, gains of a generalized system as the bandwidth and sharpness of cut-off of the random signals are varied. Considering the signals as filtered white noise, it is shown that optimization becomes difficult if the filter pole is of order greater than three, and the filter pole must be positioned carefully relative to the system´s open-loop poles, if optimization is to be achieved at all using both parameters.