A dynamic neural network designed using analytical methods produces dynamic control properties similar to an analogous classical controller

Abstract

Human balance is achieved using many concurrent control loops that combine to react to changes in environment, posture, center of mass and other factors affecting stability. Though numerous engineering models of human balance control have been tested, no methods for porting these models to a neural architecture have been established. It is our hypothesis that the analytical methods we have developed, combined with classical control techniques will provide a reasonable starting point for developing dynamic neural controllers that can reproduce classical control capabilities. In previous work, we tested this hypothesis and demonstrated that a biologically constrained neural controller that replicates human balance control characteristics is achievable. The objective of the work presented in this letter was to further understand how parameters within the neural model affect stability and correspond to expected changes predicted by classical control theory. We compare the performance between the neural and classical engineering models for bipedal balance in an inverted pendulum balance experiment. We then carry out an extended analysis on the performance of the neural controller by varying neural parameters, observing the changes in system dynamics, and comparing these changes to those predicted by the classical model. Our methods generate compact neural systems with few parameters, all of which are correlated to classical engineering parameters for a given control model. This works serves as a basis for how to port classical control problems to neural control architectures.