Parameter optimization by random search using hybrid computer techniques

Optimum selection of the parameter values for a complex dynamic system usually consists of three distinct phases: (1) a proposed system configuration is selected, in which only parameter values remain as unknowns; (2) one or more performance or cost criteria for evaluation of the system are selected; and (3) a computer technique or algorithm is chosen for adjusting the system parameters until an optimum value of the criterion function is achieved. Typical algorithms are those based on relaxation or steep descent methods. However, both of these methods are primarily suited to optimization of criterion functions with unique minima or maxima. Furthermore, they may fail to converge or may converge only very slowly if the criterion function---parameter space exhibits "ridges" or if the criterion function is only piecewise differentiable or piecewise continuous. Both of these difficulties are likely to arise in connection with nonlinear systems. This paper presents an approach to finding a global optimum by means of a modified sequential random perturbation technique implemented on a hybrid computer.