Answers are grounded in this paper and cite their sources.
In decision support systems for medical image analysis, a rule-based fuzzy classifier provides interpretability of decisions; however, its accuracy depends significantly on the parameters of the membership functions. This paper examines a fuzzy classifier in which Łukasiewicz operators and trapezoidal membership functions are approximated using a compression function based on a sigmoid function, while the rule structure is generated via a genetic algorithm. Because the compression function is differentiable, the parameters of the membership functions can be tuned using an analytical gradient. Three tuning methods—gradient-based, differential evolution, and a genetic algorithm with real-valued encoding (using BLX-α crossover)—were compared under strictly identical conditions: the same rule structure, initial parameters, objective function, and a computational budget of 3,000 evaluations. The experiments were conducted using the Iris, Wine, Pima, and WDBC datasets; the features for the WDBC dataset were derived from digital images of fine-needle breast aspirates. Gradient-based tuning across all datasets yields a statistically significantly lower mean squared error during training, reaches the final performance level of evolutionary algorithms using only 3–23% of the allocated budget, and requires 16–31% less time; the accuracy on test data is comparable across all three methods. The influence of the steepness parameter β for the squashing function was also investigated. These results justify the use of gradient-based tuning within the fuzzy component of a hybrid system for mammogram analysis.
No takes yet. Share an insight, caveat, or question.
Safin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: