Complex Adaptive System Approach to Forecasting Hurricane Tracks

摘要

Forecast hurricane tracks using a multi-model ensemble that consists of linearly combining the individual model forecasts have greatly reduced the average forecast errors when compared to individual dynamic model forecast errors. In this experiment, a multi-agent system, the Tropical Agent Forecaster (TAF), is created to fashion a 'smart' ensemble forecast. The TAF uses autonomous agents to assess the historical performance of individual models and model combinations, called predictors, and weights them based on their average error compared to the best track information. Agents continually monitor themselves and determine which predictors, for the life of the storm, perform the best in terms of the distance between forecast and best-track positions. A TAF forecast is developed using a linear combination of the highest weighted predictors. When applied to the 2004 Atlantic hurricane season, the TAF system, with a requirement to contain a minimum of three predictors, consistently outperformed the consensus forecast (CONU) at 72 and 96 hours for a homogeneous data set, although the differences were not statistically significant. But at 120 hours, the TAF system significantly decreased the average forecast errors when compared to the CONU. The multi-agent system (MAS) approach opens the door for statistically significant forecast improvement. This thesis was submitted in partial fulfillment of the requirements for the degree of Master of Science in Computer Science.