Determining fuzzy integral densities using a genetic algorithm for pattern recognition

By utilizing information provided from different sources, data fusion has been a very effective method to achieve good performance in many applications, such as pattern recognition and decision making. The fuzzy integral is one of the many ways to combine those information sources. Successful applications have shown that, if used properly, the fuzzy integral can be a powerful tool in dealing with data fusion problems. There is however, a key issue unsolved in the application of fuzzy integrals - the determination of density values which determine the fuzzy measure used in the fusion process. Although the densities can be interpreted as the relative importance of information sources to be combined, how to calculate them remains a problem. Since the performance of fuzzy integral largely depends on the densities, density selection is critical. In this paper, a genetic algorithm (GA) has been used to search for an optimal set of density values. This method was applied to a handwritten digits recognition problem. Outputs of six neural network classifiers were combined using the fuzzy integral whose densities were obtained from the genetic algorithm. The experiment showed that fuzzy integral using densities calculated from GA outperformed that using fixed densities, those obtained from averaging of the classifiers' outputs as well as the results of individual neural network classifiers.