Abstract
Multilayer fuzzy connective-based hierarchical aggregation networks simulate the decision-making processes performed by humans, and the results can be interpreted as a set of rules. Identifying the relative importance of the inputs helps to identify redundancies that do not contribute to the decisionmaking process. However, a gradient-based learning approach tends to generate local solutions, and requires the aggregation function to be continuous and differentiable. This study proposes a GA-based learning approach to identify the connective parameters, exploiting the global exploration ability of GAs to improve the quality of solutions. This approach does not require gradient information, making it applicable to both differentiable and nondifferentiable aggregation functions. Statistical analysis of the experimental results confirms that the proposed approach outperforms the gradient-based learning approach, generating more accurate estimates for both generalized mean and gamma operators. © 2011 IEEE.