Implementing the Taguchi-Statistical Learning-DEAR Methodology in a MultiCriteria Decision Making Approach to Balance Trade-offs in Evolutionary Algorithm Performance

Abstract

This work presents a new approach to improving the efficiency of Evolutionary Algorithms (EA) in extremely noisy function landscapes. Statistical Learning, Taguchi analysis and normalization, and Data Envelopment Analysis based Ranking (DEAR) are used to provide a hybrid technique that provides a complete framework for EA parameter adjustment. The study examines the effects of four important EA parameters on function yield and computing time: convergence, mutation rate, population size, and random seed. Taguchi analysis and normalization is used to generate an efficient experimental design that covers different combinations of parameter values, allowing a methodical exploration of the parameter space. Subsequently, the DEAR approach is employed to prioritize each set of parameters according to certain optimization criteria. To further complicate matters, the optimization goals and EA parameters are both modeled using Statistical Learning approaches. There has been a lot of testing with noisy functions of artificial landscapes with three different types: single-peak, curved-ridge, and multi-peak. Assuming a normally distributed distribution with a mean of 0 and standard deviations of 0.05 and 0.2, noise presents practical obstacles to optimization. When compared to more traditional approaches of parameter tuning, the suggested hybrid strategy clearly outperforms the competition in terms of computing time, function yield, and mean and standard deviation of both metrics. The technique shows improved resilience and adaptability across varied noisy environments and more successfully finds optimal parameter configurations, according to the results. By demonstrating its flexibility to meet evolving optimization needs, sensitivity assessments provide more evidence of the suggested methodology's dependability. Finally, the research presents a state-of-the-art hybrid method for tweaking evolutionary algorithm parameters, which considerably improves upon previous efforts. In particular, when it comes to dealing with complicated and noisy optimization scenarios, the offered technique stands out due to its capacity to continuously produce greater performance, making a vital addition to the optimization community.