optimization of devices innehaallande several components, saerskilt electronic circuits saasom filter optimization of exchange in the manufacture of devices innehaallande several components bestaemning of viktskoefficienter in neutral naet

摘要

Optimization by means of Gaussian adaptation is proposed for example for arrangements including a number of components, for manufacturing efficiency, for regulator circuits, for weight coefficients in neural networks. Generation of a Gaussian-distributed random sample g with a mean value 0 and fixed variance greater than 1 is carried out, after which this random sample is multiplied by a square matrix W, which change during the course of adaptation, so that y = Wg is obtained. A random sample x of the desired component values, weight coefficients etc. is obtained from calculation of the expression x = m + y, where m is a variable mean value for the component values, weight coefficients etc. For this random sample of x, the value of a quality or criterion function f is then calculated. If this has a sufficiently good value, the random sample of x is accepted and new values for the mean value and the square matrix are calculated by means of the expressions m = (1-a)m + ax and W = (1-b)W + bygT. Generation of a new Gaussian-distributed random number g is then carried out and the above procedure is repeated until a sufficiently good value of the criterion function has been obtained. In this way, very simple and rapid optimization is obtained for cases where the criterion function can have an arbitrary appearance, for example it may be discrete, non-derivable etc. IMAGE