Yerkes-Dodson Law in agents' training

Abstract

Well known Yerkes-Dodson Law (YDL) claims that medium intensity stimulation encourages fastest learning. Mostly experimenters explained YDL by sequential action of two different processes. We show that YDL can be elucidated even with such simple model as nonlinear single layer perceptron and gradient descent training where differences between desired outputs values are associated with stimulation strength. Non-linear nature of curves "a number of iterations is a function of stimulation" is caused by smoothly bounded nonlinearities of the perceptron's activation function and a difference in desired outputs. © Springer-Verlag Berlin Heidelberg 2003.