Strategies for Better System Identification

该项目将研究各种方法，以提高人工神经网络（ANN）学习分类或一般数学函数的能力。它将研究人工神经网络的替代功能形式，建立在PI先前的径向基本函数和b样条的工作基础上。它将基于随机梯度学习理论的思想，研究适应学习率的方法。它将研究除了通常的平方误差函数之外的其他误差函数，旨在产生更好的泛化能力。这样的可选误差函数，旨在产生更好的泛化能力。这种可选的误差函数也可用于确定人工神经网络的拓扑结构。

This project will investigate a variety of approaches, to improve the ability of artificial neural networks (ANN) to learn classifications or general mathematical functions. It will investigate alternative functional forms for ANNs, building on the PI's prior work with radial basic functions and B-splines. It will investigate methods for adapting learning rates, based on ideas about stochastic gradient learning theory. It will investigate alternative error functions, beyond the usual square error function, designed to yield better generalization power. Such alterative error function, designed to yield better generalization power. Such alternative error functions may also be used in determining the topology of an ANN.