Combining Dialectical Optimization and Gradient Descent Methods for Improving the Accuracy of Straight Line Segment Classifiers
Rosario Medina-Rodríguez, Ronaldo F. Hashimoto · 2011
A recent published pattern recognition technique called Straight Line Segment (SLS) uses two sets of straight line segments to classify a set of points from two different classes and it is based on distances between hese points and each set of straight line segments. It has been demonstrated that, using this technique, it is possible to generate classifiers which can reach high accuracy rates for supervised pattern classification. However, a critical issue in this technique is to find the optimal positions of the straight line segments given a training data set. This paper proposes a combining method of the dialectical optimization method (DOM) and the gradient descent technique for solving this optimization problem. The main advantage of DOM, such as any evolutionary algorithm, is the capability of escaping from local optimum by multi-point stochastic searching. On the other hand, the strength of gradient descent method is the ability of finding local optimum by pointing the direction that maximizes the objective function. Our hybrid method combines the main characteristics of these two methods. We have applied our combining approach to several data sets obtained from artificial distributions and UCI databases. These experiments show that the proposed algorithm in most cases has higher classification rates with respect to single gradient descent method and the combination of gradient descent with genetic algorithms.