Population Generation Methods for Metaheuristic Algorithms Used to Construct Compact Fuzzy Classifiers of Medical Data

Fuzzy classifiers differ from other machine learning algorithms in their ability to interpret the inference process, which is especially important in high responsibility subject areas such as medicine. The membership functions of fuzzy terms and the rule base are easy to visualize, so it is not difficult for the user to understand why a particular result was obtained. However, the interpretability of the model suffers if the resulting model is highly complex, when the classifier has more than a dozen of high-length rules. Balancing accuracy and complexity in fuzzy classifiers is a nontrivial task. This article, the first in a series about constructing compact classifiers for medical data, addresses the problem of maximizing accuracy with as few rules as possible using metaheuristic algorithms. Using metaheuristics to optimize fuzzy rules allows a more accurate representation of the subject domain, which has a positive effect on classification accuracy. To increase the efficiency of population metaheuristics, it is important to use an appropriate method for a particular algorithm to form optimization starting points. The paper investigates the effect of using different population identification methods for two metaheuristics – the swallow swarm algorithm and the hybrid of the gravitational search algorithm and the shuffled leaping frogs algorithm.