Predicting Nonlinear Time Series
Abstract
Predicting future values of a time series has many practical uses in real-time signal processing and understanding. This thesis implements an Adaptive Time Delay Neural Network (ATNN) capable of user-defined degeneration to the more common Time Delay Neural Network (TDNN). Time delays along axons or at the synapses, which vary in biological systems, motivate this research. The ATNN/TDNN test results and time series prediction capabilities are compared to those of the Real-Time Recurrent Learning (RTRL) algorithm. To show the advantages and disadvantages of using TDNN and ATNN for prediction versus the RTRL, the networks were applied to two problems: incommensurate sum of sine waves and financial time series. These data sets represent examples of nonlinear data with known and unknown mathematical functions, respectively. Although the RTRL predicted better than the ATNN for a known predictable function, this ATNN approach proved competitive in determining the direction of the future values for this function and outperforms the RTRL on the more difficult prediction task. The ATNN program, developed in C++ with an object-oriented framework, also takes much less computation time than the RTRL during training.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1993
- Accession Number
- ADA274051
Entities
People
- James C. Gainey Jr
Organizations
- Air Force Institute of Technology