Towards Human Control Strategy Learning: Neural Network Approach with Variable Activities Functions

Abstract

Human beings epitomize the concept of "intelligent control". Despite its apparent computational advantage over humans, no machine or computer has come close to achieving the level of sensor-based control of which humans are capable. Thus, there is a clear need to develop computational methods which can abstract human skill based on sensory feedback. Neural networks offer one such method with their ability to map complex nonlinear functions. This paper is divided into two parts. First, we examine the problem of approximating continuous functions, as is required for dynamic system identification and control. We discuss how the requirements of continuous function approximation differ substantially for those of discrete function approximation. To meet these requirements, we propose to use the cascade two learning architecture, which dynamically adjusts the size of the neural network as part of the learning process. As such, we propose different hidden units to have variable activation functions, leading to faster learning, as well as better function approximation. Second, we apply these methods towards the problem of control system identification, and more specifically, to the problem of identifying and modeling human control strategy. We demonstrate the feasibility of the proposed method in human control strategy or skill learning, and address issues for potentially exciting research in the future. This approach can play a significant role in the development of intelligent machines based on human skill learning, as well as in the intelligent and harmonious and coordination of humans and robots.