Dynamic Allocation in Neural Networks for Adaptive Controllers

Abstract

Dynamic allocation in Neural Networks is the process of strategic addition of nodes during the evolution of a feature map. As the trend of using growing neural networks is rising in adaptive controller applications it is important to understand the robustness of the process of dynamic allocation in neural networks. In this paper we analyze the robustness of the process of dynamic allocation that are commonly utilized in growing neural networks to varying, and non-stationary input data. The analysis indicates that dynamic allocation in growing neural networks is not fully robust if based solely on the information from resource values or connectivity structure of the nodes. Based on the observations made, we propose a data-driven dynamic allocation algorithm that is useful for growing neural networks used in adaptive controller applications. The advantage of the proposed algorithm is that it allows neural networks to localize the information represented in the input data while ensuring that the overall topology of the data is preserved. Experimental results are presented to demonstrate using high-dimensional, multivariate data obtained from an adaptive flight controller simulator. The analytical and experimental results affirm the robustness and establish the precedence of the developed dynamic allocation algorithm for adaptive controller applications. We investigate the process of dynamic allocation in the Dynamic Cell Structures neural network algorithm, a representative growing neural network used for on-line learning in adaptive controllers, but the approach presented is applicable to any growing neural network where node insertion is performed in order to improve data modeling.