Approximation of the inverse kinematics of an industrial robot by DEFAnet

A deterministic network concept that is capable of approximating arbitrary continuous functions with any desired accuracy is presented. A DEFAnet is a four-layered feedforward network. The outputs of each neuron are monotonous functions of the sum of the neuron´s inputs weighted with the synaptic strengths. The DEFAnet approach has been used to approximate part of the inverse kinematics of an industrial robot with six degrees of freedom. It is shown that both calculation and learning may yield reasonable approximations. The accuracy attainable with a given network size can be considerably improved by adjusting a set of smoothing parameters. In addition, the accuracy improves better than proportionally to the number of neurons