SERI: Stagnation-Based Extinction and Re-initialization Operator for Enhanced Evolutionary Optimization

Evolutionary algorithms are continually evolving and remain widely used for tackling complex optimization problems including neural processing architectures and hyperparameter tuning. However, as they typically rely on operators such as selection, crossover, and mutation, they are often prone to stagnation, loss of diversity, and premature convergence, which can hinder their ability to reach the global optimum. Based upon the recent work inspired by biological evolution and building upon Peircean Evolutionary Algorithm (PEA), which integrates a philosophically-inspired triadic model and highlights the value of introducing additional mechanisms to overcome these limitations, our paper proposes a novel Stagnation-Based Extinction and Re-initialization (SERI) operator. The design of this operator aims to actively detect stagnant individuals in the population and, through re-initialization, inject fresh diversity to sustain exploration and enhance overall performance. Experimental results on various benchmark test functions, including the Competition on Evolutionary Computation 2022 (CEC’22) suite, demonstrate the effectiveness of our proposed approach. The comparative analysis highlights the potential of SERI to improve the performance of population-based optimization algorithms. Additionally, we have tested SERI on fine tuning hyperparameters and CNN architectures.