Research Initiation Award: Predicting and Characterizing Noisy Time Series

[309786] Weigend在过去十年中，在预测和理解时间序列方面取得了显著的进展。时间序列分析曾经是由线性系统理论形成的，现在有可能认识到一个明显复杂的时间序列是由一个低维非线性系统产生的，表征其基本属性，并建立一个可用于预测的模型。在另一个极端，现在有一个更丰富的框架来设计算法，比如神经网络，它可以学习和适应时间序列中没有简单起源的结构。本文主要研究了三个问题：如何估计时间序列预测的精度，如何处理有噪声和混沌的数据集，以及如何通过分析预测模型来表征时间序列系统。为了回答这些问题而开发的工具结合了连接主义和动力系统理论的最新进展。他们将根据现实世界的数据进行评估，这些数据由不同的团体提交，供圣达菲研究所主办的时间序列预测和分析竞赛审议。* * *

9309786 Weigend The progress in the last decade in predicting and understanding time series has been remarkable. Where once time series analysis was shaped by linear systems theory, it is now possible to recognize when an apparently complicated time series has been produced by a low- dimensional nonlinear system, characterize its essential properties, and build a model that can be used for prediction. At the opposite extreme, there is now a much richer framework for designing algorithms such as neural networks that can learn and adapt to the structure in time series that do not have a simple origin. This research addresses the following three questions: How to estimate the accuracy in time series prediction, how to deal with data sets that are noisy and chaotic, and how to characterize the system that temporal sequence by analyzing the predictive model. The tools to be developed in response to these questions combine recent advances from connectionism and dynamical systems theory. They will be evaluated on real-world data, submitted by various groups for consideration at the Time Series Prediction and Analysis Competition that was held under the auspices of the Santa Fe Institute. ***