Researchers have developed an enhanced metaheuristic framework, termed the Elite-Guided Adaptive Phase Switching Dual-Leader Piranha Optimisation Algorithm (EG-APS-DLPOA), to tackle core limitations in global numerical search and automatic data clustering. Proposed by Subrat Kumar Nayak, Biswa Ranjan Senapati, and Tribhuvan Singh, the framework builds upon the baseline Piranha Optimisation Algorithm by introducing three dedicated structural mechanisms. The design directly addresses three prevalent challenges in metaheuristic computing: balancing exploration with exploitation, preventing early search convergence, and maintaining optimization performance as problem dimensionality increases.
To overcome these challenges, the authors incorporated adaptive phase switching, dual-leader guidance, and elite-agent stratification into the original search procedure. Under adaptive phase switching, a decaying probability governs the optimization trajectory rather than a predetermined schedule, prioritizing broad exploration during earlier iterations and transitioning toward focused exploitation in later stages. Dual-leader guidance supplies directional control during the exploratory phase by simultaneously leveraging the global best solution and the mean position calculated from elite agents. Complementing these adjustments, elite-agent stratification isolates the top 20 percent of candidate agents as an elite tier, mandating that these elite individuals exclusively exploit the global best position without engaging in exploratory movement.
To isolate and evaluate the individual contribution of each component, the investigators conducted a sequential ablation study across four incremental configurations: the original Piranha Optimisation Algorithm, a dual-leader variant, an adaptive phase-switching dual-leader version, and the full elite-guided framework. Experimental validation encompassed 38 mathematical benchmark functions divided across two distinct test environments. An 18-function classical suite served as the initial baseline, evaluated at dimensions of 30 and 50 over 30 independent runs. In parallel, a 20-function test suite drawn from the CEC2017 benchmarks provided rigorous validation on complex, shifted, rotated, and hybrid landscapes across dimensions of 30, 50, and 100.
Across the 18 classical benchmark functions, the full algorithm ranked first at 30 dimensions and second behind the Whale Optimization Algorithm at 50 dimensions. Statistical assessments indicated that the performance of the proposed algorithm was statistically equivalent to the Whale Optimization Algorithm, the Grey Wolf Optimizer, and its own intermediate ablation variants on this suite. The comparative records showed that the Whale Optimization Algorithm secured top outcomes on the majority of separable functions within the classical test suite, indicating that the new framework’s comparative advantages are context-dependent.
The framework demonstrated its most distinct performance advantages on the CEC2017 benchmark suite, securing the first-place overall rank across all three evaluated dimensions. When tested at 30 dimensions, the algorithm performed significantly better than 8 of 10 competing algorithms, extending that advantage to 9 of 10 competitors at 50 dimensions, and outperforming all 10 rival algorithms at 100 dimensions. These empirical outcomes highlighted that the framework’s superior performance is clearest on rotated, hybrid, and non-separable landscape topologies. Conversely, on simpler separable functions, the algorithm proved competitive with the Whale Optimization Algorithm and the Grey Wolf Optimizer without demonstrating clear superiority.
Beyond mathematical function benchmarks, the authors evaluated the algorithm’s utility in unsupervised pattern recognition using 11 UCI machine learning clustering datasets. In these automatic clustering tasks, the framework secured a second-place overall ranking while recording the highest classification Accuracy and the highest Adjusted Rand Index among all assessed algorithms. Performance analyses also revealed that the specific benefit of elite stratification expanded as dimensional complexity increased, providing greater relative improvements at 50 and 100 dimensions than at 30 dimensions on the CEC2017 suite.
Subject of Research: Metaheuristic global optimization and automatic clustering algorithms
Article Title: An Adaptive Dual-Leader Optimization Framework with Elite Stratification for Global Optimization and Automatic Clustering
Article References: Nayak, S. K., Senapati, B. R., & Singh, T. (2026). An Adaptive Dual-Leader Optimization Framework with Elite Stratification for Global Optimization and Automatic Clustering. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01684-3
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01684-3
Keywords: EG-APS-DLPOA, Piranha Optimisation Algorithm, Metaheuristic algorithms, Global optimization, Automatic clustering, Dual-leader guidance, Adaptive phase switching, Elite-agent stratification, Exploration and exploitation, CEC2017 benchmark suite, High-dimensional optimization, UCI clustering datasets, Adjusted Rand Index, Non-separable functions, Hybrid benchmark functions, Whale Optimization Algorithm, Grey Wolf Optimizer
Cite Scienmag News
Denise Maddox. (November 18, 2026). Elite-Guided Piranha Optimizer Outperforms Competing Algorithms on Complex Benchmark Functions. Scienmag. https://scienmag.com/elite-guided-piranha-optimizer-outperforms-competing-algorithms-on-complex-benchmark-functions/
Denise Maddox. "Elite-Guided Piranha Optimizer Outperforms Competing Algorithms on Complex Benchmark Functions." Scienmag, 18 November 2026, https://scienmag.com/elite-guided-piranha-optimizer-outperforms-competing-algorithms-on-complex-benchmark-functions/. Accessed 18 November 2026.
Denise Maddox. "Elite-Guided Piranha Optimizer Outperforms Competing Algorithms on Complex Benchmark Functions." Scienmag. November 18, 2026. https://scienmag.com/elite-guided-piranha-optimizer-outperforms-competing-algorithms-on-complex-benchmark-functions/








