Evaluating Adaptive Oscillatory Neural Network Controllers using a Simple Vehicle Model

Abstract

Parameters values estimation is a non trivial task, particularly when it applies to the problem of an oscillatory Recurrent Neural Network (RNN) controller. Also, the link between the controller and the physical body is crucial in adaptive processes studies, but most of the bodies in the current literature are too complex to clearly analyze the possibilities of adaptability of the controller in interaction with the body. In this paper, we use in comparison a simple mechanical system called the Roller Racer, to be able to test learning strategies with oscillatory RNN controllers. We briefly present the Roller Racer model, and build two possible architectures of a RNN controller for it. The parameters values are estimated with a variation of the classical teacher forcing learning algorithm, which we extend to "teach" signals with a continuous component.