Adaptation of Learning Parameters for Choquet IntegralAgent Network by Using Genetic Algorithms

摘要

Choquet Integral Agent Network (CHIAN) is proposed as a method realizing flexibleinformation fusion which is constructed by using fuzzy measure and Choquet integrals. In caseof multi-layered network structure, CHIAN can employ back-propagation algorithms-likeconcept for learning process. However, the back-propagation methods have some limitationssuch as trapping at local minima and network paralysis. Due to genetic algorithms (GA) mecha-nism, it has the characteristics of hill climbing, and thus can overcome the difficulty of trappingat local minima; consequently it might reduce network paralysis. This paper aims at proposingto tune CHIAN learning parameters, i.e., learning rate and momentum coefficient by geneticalgorithms for improving CHIAN as classifier, pattern recognition, and information fusion. Theresults show that the network evolved GA requires fewer training cycles than the networkwhich the learning parameters are intuitively given.