Research Information System
Applied Soft Computing, vol.177, 2025 (SCI-Expanded)
Fuzzy cognitive maps (FCMs) are well known for their ability to learn and analyze the dynamic behavior of complex systems, but their effectiveness is often hindered by classical reasoning mechanisms, which are rooted in control systems. Regardless of the initial conditions, they tend to converge to a steady state, leading to poor predictive performance in learning FCMs. To address this issue, this study proposes a novel reasoning mechanism tailored for error-driven learning in FCMs. By incorporating indirect relationships into the reasoning mechanism, the proposed method provides a more comprehensive assessment of each concept's influence within the system. This allows for a more effective response to initial conditions and improves prediction accuracy. The proposed method employs the Particle Swarm Optimization (PSO) algorithm to learn relationships and is evaluated through extensive experimental studies using three real-world datasets (scenarios). The experiments compare the proposed method with classical and recently developed reasoning mechanisms by systematically varying normalization techniques, fitness functions, and computational methods. The findings demonstrate that the proposed method consistently outperforms classical methods in prediction accuracy across multiple scenarios. For instance, in one scenario, the proposed method achieves up to 49 % lower out-of-sample error and correctly predicts 83 % of concepts, outperforming sigmoid and hyperbolic tangent-based methods, which achieve 51 % and 50 % accuracy, respectively. Notably, Heaviside emerges as the most effective fitness function across all experimental setups, delivering the lowest prediction errors and improving accuracy by 12–35 % using the proposed reasoning mechanism. Comparisons with exponential normalized reasoning method reveal that the proposed method achieves superior prediction accuracy in 90 % of test cases, demonstrating statistically significant improvements in most setups. Overall, the results demonstrate the superiority of the proposed method in learning FCMs, offering enhanced prediction accuracy and generalization across diverse scenarios.