- author = "Andrew John Novobilski",
- title = "Forecasting of Time Series Data Using Naturally Selected {Bayesian} Networks",
- school = "University of Texas at Arlington",
- year = "2000",
- address = "USA",
- month = may,
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://genealogy.math.ndsu.nodak.edu/id.php?id=60699",
-
URL = "http://search.proquest.com/docview/304675953",
- size = "112 pages",
-
abstract = "Time series forecasting is a well-established area of
study supported by various methods for predicting a
future value from historical information. In addition
to continued research into methods originating from
numerical analysis, there is a growing interest in
methods that enhance numerical techniques by providing
a network-based infrastructure for describing the
organization of the series being forecast. A common
factor shared by network-based techniques such as
neural networks and Bayesian networks is their reliance
upon cognitive science to provide the inspiration for
the computational model the network represents. One
common problem they share is the automated
identification of the appropriate network structure
that is used to model the time series being
forecast.
This dissertation addresses the problem of forecast model identification by extending the use of cognitive science techniques into a framework for the natural selection of Bayesian networks. This framework consists of a network model that has two components. The first component is a genetic programming based representation that decomposes into two parts; decision making and information gathering. The decision making part is represented as a Modified Naive Bayesian classifier. The information gathering part is represented as an attribute generator that classifies input values according to numerical categories. The search process for identifying the network components is done using natural selection as implemented using genetic programming. The second framework component uses the resultant Bayesian model to forecast future values of a given time series by using the maximum a posteriori (MAP) hypothesis generated by querying the Bayesian network.
Using the framework for naturally selecting Bayesian networks, experimental results are presented for both synthesized time series and data sets drawn from the daily values of individual stocks traded on the NASDAQ and NYSE. Using a Random Walk as the benchmark forecast, results from three variations of the naturally selected Bayesian network forecaster with single variable inputs are compared to results obtained using a neural network. Additional information detailing network structure and forecasting accuracy is presented for the best models describing each time series forecasted.",
- notes = "UMI 9975303 supervised by Farhad Ahani Kamangar",
