Central pattern generator and its learning via simultaneous perturbation method

Abstract

In this paper, we propose models and learning schemes of central pattern generator(CPG). The CPG models consist of plural neural oscillators which generate simple waves. Combining the neural oscillators, the model can generate complicated waveforms. In order for the CPG to generate desired wave, it is important and essential to present a suitable learning scheme. In this paper, learning schemes using the simultaneous perturbation optimization method is introduced. This learning scheme utilizes only output of the CPG. Therefore, unlike the ordinary back-propagation learning rule, the proposed learning scheme is easily applicable to the CPG models. Moreover, complex-valued CPG is also proposed. In the CPG, inputs, outputs and the other variables are basically complex numbers. Learning scheme based on the simultaneous perturbation method is also introduced. Walking motion control for humanoid robot is considered as an example. The CPG could learn and control ten joint angles of the robot for walking and stepping motion patterns. Moreover, three different desired waveforms in real part and imaginary part are considered for the proposed complex-valued CPG.