Investigation of Reference Sample Reduction Methods for Ensemble Output with Fuzzy Logic-Based Systems

Abstract

One of the main methods in data reduction processes is the instance selection method. Reducing the dataset has two main objectives: reducing the requirements for computing resources, and the time for processing the learning task. The paper studies the problem of reducing the size of a reference set of points (reference sample) in collective decision making. In the paper, the reference sample refers to the sample that is used in ensemble output based on fuzzy logic system. The fuzzy controller makes a decision which agent should be used for each point from a test set. The nearest point from the reference sample is determined for any point from a test set. Depending on the distance to the object from a test set and the successfulness of the algorithm on this object, the confidence of the algorithm on this test point is determined. Also, it is proposed to apply the instance selection to choose instances for the reference set from the training set when solving regression problems based on such methods as genetic algorithms (GA), the k-means clustering algorithm, and the random instance selection (RIS). Computational experiments show that effective instance selection in the reference set can significantly reduce the computational costs while maintaining the accuracy of the result.