A study of generalization ability of neural network for manipulator inverse kinematics

The authors propose a method to determine the optimal unit number in the hidden layer of a feedforward-type neural network. The generalization ability of the three-layer neural network is influenced by the number of the hidden units. In this method, the relationship between the hidden and output layer is formulated by the multiple regression model, and the unit number in the hidden layer which minimizes the AIC (Akaike´s information criterion) is adopted as the optimal unit number for the not training set. The effectiveness of the proposed method was confirmed from the simulation results for the inverse kinematics problem of a two-link robot manipulator