Recurrent Neural Networks and Motor Programs

摘要

It is well known that biological control systems have a hierarchical design, consisting of brain, central pattern generators (CPGs) and actuators. The presence of the CPG means that descending control signals from the brain tend to be of a simple form. In this paper, this general design principle is implemented in a neural network framework. CPGs are implemented as fully recurrent nets, with Distal Learning used to estimate error derivatives for a controller. This ideas are tested on a simulated robotic arm.