Parameter extraction of microwave transistors using a hybrid gradient descent and tree annealing approach

Tree annealing is a robust optimization scheme which can be used to find the valleys of an error surface. The problem of entrapment in local minima is not a factor with this type of optimization, but it is much slower than gradient-based techniques. The method presented here attempts to take advantage of the speed of gradient-based methods and of the efficient pseudorandom searching abilities of tree annealing. The result is a technique which behaves as a directed multistart gradient method. All minima encountered during optimization are recorded, thus providing alternatives in case of a nonphysical final solution. The technique is used in the extraction of a modified Materka-Kacprzak model of a GaAs MESFET.<
>