SBIR-STTR Award

Time series prediction techniques are central to a number of broad application areas critical to both defense and civilian industry, including weather forecasting, economic forecast and a variety of signal processing and control applications. Many specific application problems involve intrinsic nonlinearity which the well-established linear prediction methods (e.g., Arma, Arima, FIR and IIR filters, Kalman filters, LMS filters, etc.) are unable to capture effectively. Nonlinear prediction models using neural networks (such as multilayer perceptrons and radial basis functions) offer the possibility of substantially improved performance over conventional linear models in predicting the behavior of nonlinear systems. A number of critical issues must be addressed in the development of appropriate nonlinear prediction models, such as the selection of variables, architecture selection, architecture pruning, the estimation of prediction risk (generalization performance). Phase I of the project will focus on the development of statistically robust and computationally efficient algorithms and techniques for constructing neural network time series predictors. Weather forecasting will serve as a test problem for Phase I. Phase II of the project will use the algorithms techniques developed in Phase I to solve one or more large scale forecasting problems of direct interest to defense or commercial clients.Anticipated benefits/potential applications:Developing statistically robust methods for nonlinear neural network time series prediction systems will allow NPS to solve a number of problems of practical interest, including short-term, localized forecasting of weather.

Time series analysis and prediction is very difficult for series that have substantial noise, nonstationarity, or nonlinearity. Nonlinear Prediction Systems' Phase I research demonstrated that state-of-the-art neural network algorithms can yield superior prediction performance over standard time series methods for a variety of hard problems. NPS proposes to significantly refine and extend this work during its Phase II R&D. The main fundamental technical issues to be addressed are the choice and fine-tuning of methods for: input variable selection, time series representations, neural network model specification and selection, the bias/variance and noise/nonstationarity tradeoffs, regularization, minimizing prediction risk, specialized time series modeling methods, and dynamic learning. A software infrastructure will be built for rapidly developing statistically-robust neural network time series solutions for target applications. The technology will be applied to modeling a number of very important and very difficult time series: the major U.S. macroeconomic series and several major financial and strategic commodities markets (strategic metals and energy.) The neural network time series models will go beyond simple forecasting models will include decision making and trading systems. Improved forecasting systems for the macroeconomy and forecasting and trading systems for the financial and strategic commodity markets are important for U.S. economic growth and U. S. economic security. Anticipated

Benefits:
The results of the Phase II R&D will be: (1) a technology and software infrastructure for rapidly developing statistically-robust neural network based time series models, and (2) reliable and commercially useful forecasting systems for the major U.S. macroeconomic time series and forecasting and trading systems for several major financial and strategic commodities markets. Potential users of the technology and the target application solutions include various government agencies and commercial enterprises.