Agent-Based Cluster Analysis of Tropical Cyclone Tracks in the Western North Pacific

Abstract

The clustering model integrating Finite Mixture Model (FMM) and classical Expectation-Maximum (EM) algorithm has been applied to tropical cyclone (TC) tracks during the last decade. However, the efficiency of classical EM algorithm is insufficiently good and the robustness of the model is not verified. Besides, it is inconvenient for users to manually choose the parameters for the cluster analysis. In order to improve the efficiency of classical EM algorithm, the "Lazy-Ψα 2" EM is proposed by integrating Lazy EM algorithm and Ψα 2 algorithm. Sensitivity analysis is conducted to ensure the insensitivity of the clustering model to the amount of data set. The cluster analysis is implemented on an agent-based framework by which the tool can automatically choose the parameters by evaluating the clustering performance. TC tracks in western North Pacific from 1949 to 2006 are classified into 12 clusters by the probabilistic clustering model that is solved by "Lazy-Ψα 2" EM algorithm. The log-likelihood is taken as the performance indicator. Elaborate comparisons are made between the present cluster analysis and other cluster analyses related to TC tracks.