Seleção automática de parâmetros iniciais do algoritmo K-segmentos com teaching-learning-based optimization

Abstract

Being a nonlinear generalization of principal component analysis, the principal curves technique is a robust tool for data analysis and classification. In pattern recognition one of the most popular algorithms to build Principal Curves is the k-segments. This algorithm presents good results and excellent applicability due to its guaranteed convergence and robustness. However, its use depends on user-defined parameters. This work presents an automatic selection technique of the quantity and length of segments of the k-segment algorithm using the TLBO (Teaching-Learning-Based Optimization) metaheuristic. A cost function is used that takes into account the length of the curve and the distance of the events to the segments where they are projected. Experimental tests made with synthetic databases are presented to demonstrate the efficiency of the proposed method in representation problems.