On Convergence of Neural Networks Learning in the Presence of Constantly Changing Data and its Applications to Safety Monitors

摘要

The problem considered in this paper is that of a feed forward neural network learning in the presence of constantly changing data. We show that under some very general conditions learning can be adapted to assure that the solution stays in proximity to certain equilibrium conditions. We then suggest an algorithm for monitoring neural net performance in the aircraft control applications. This paper addresses the problem of real time assessment or evaluation of the performance of an adaptive neural network controller. Complex control systems require robust controllers that handle a large variety of operating conditions. In these circumstances controllers must be adaptive or robust to accommodate inaccurate plant models and changes in the plant dynamics. One of the major difficulties related to adaptive control is proof of stability of the update mechanism. In the case where a conventional controller is available it is convenient to toggle between the conventional and adaptive based on controller performance. This paper presents a proposed approach to evaluate the performance of an adaptive sigrna pi and a single hidden layer neural network controller.