CISE Postdoctoral Program: Robust Forecasting with Neural Networks

在该项目中，将开发用于预测时间序列的鲁棒人工神经网络（ANN）学习算法和工具，并将其具体应用于金融和经济时间序列。由于噪声的非正态分布和异常值的存在，以及噪声和信号的非平稳性，这些算法和工具难以预测。因此，它们为鲁棒非线性建模技术提供了很好的基准问题。在健壮模型中，将纳入健壮误差度量，以优雅地处理数据中的异常值。同时，提出了一种利用代价函数梯度的二阶统计量在训练过程中自动正则化（约束）神经网络模型的新方法。这阻止了网络对噪声的适应，从而产生了鲁棒模型。在学习算法方面，将研究一种新的模拟退火型学习算法在循环网络中的应用。基于以往对多隐层前馈网络的研究经验，相信该算法将提高递归网络的性能。在变量选择方面，将遗传算法（GA）与人工神经网络模型相结合，以选择最优变量子集来解决问题。将GAs与一种新开发的统计检验（即所谓的delta检验）结合起来，成为一种更强大的工具，用于确定最优变量子集。＊＊＊

9503968 Moody In this project robust artificial neural network (ANN) learning algorithms and tools for forecasting time series, with specific applications to financial and economic time series will be developed. These algorithms and tools are difficult to predict due to the non-normal distribution of the noise and the presence of outliers, as well as the nonstationarity of both noise and signals. They thus make excellent benchmark problems for robust nonlinear modeling techniques. In robust models, robust error measures for gracefully handling outliers in the data will be incorporated. Also, a new method for automatically regularizing (constraining) the ANN model during training by using second order statistics of the cost function gradient. This prevents the network from adapting to noise and thus produces robust models. In the area of learning algorithms, the application of a new simulated annealing type learning algorithm to recurrent networks will be investigated. It is believed that, based upon past experience with feedforward networks with many hidden layers, this algorithm will improve the performance of recurrent networks. In variable selection, a novel use of genetic algorithms (GA) in combination with ANN models to select optimal subsets of variables for solving a problem will be pursued. Combining GAs with a newly developed statistical test, the so-called delta-test, into a more powerful tool for determining the optimal variable subset will also be done. ***