An Adaptive Dual-Leader Optimization Framework with Elite Stratification for Global Optimization and Automatic Clustering

Metaheuristic algorithms face three common problems. It is hard to balance exploration and exploitation. Search can converge too early. Performance often drops as the dimension grows. This paper proposes EG-APS-DLPOA, the Elite-Guided Adaptive Phase Switching Dual-Leader Piranha Optimisation Algorithm. It is built on the Piranha Optimisation Algorithm (POA) and adds three mechanisms to it. The first mechanism is adaptive phase switching. A decaying probability controls the search. Early iterations favour broad exploration, while later iterations favour focused exploitation, with no fixed schedule. The second is dual-leader guidance. It adds direction to the exploration step, using the global best position and the mean position of elite agents. The third one is elite-agent stratification. The top 20% of agents are marked elite. Elite agents always exploit the global best and never explore. An ablation study tests each mechanism alone, in order: POA, DL-POA, APS-DLPOA, then EG-APS-DLPOA. Each step adds one mechanism, isolating its effect. The algorithm is tested on 38 benchmark functions. Eighteen form the main suite, tested at D = 30 and D = 50 over 30 runs each. The other 20 come from the CEC2017 suite, giving an independent validation test at D = 30, D = 50 and D = 100, over 30 runs each on harder, shifted, rotated, and hybrid functions. On the classical suite, EG-APS-DLPOA ranks first at D = 30 and second behind WOA at D = 50. It is statistically equivalent to WOA, GWO and its own ablation variants, and WOA wins most of the separable functions. On the CEC2017 suite , EG-APS-DLPOA ranks first at all three dimensions. It is significantly better than 8 of 10 competitors at D = 30 , 9 of 10 at D = 50 and all 10 at D = 100. Its advantage is therefore clearest on rotated, non-separable and hybrid landscapes. On separable functions, it is competitive with WOA and GWO, but not superior to them. The algorithm is also tested on 11 UCI clustering datasets. It ranks second overall, with the best Accuracy and ARI of all algorithms tested. On CEC2017, the benefit of elite stratification is larger at D = 50 and D = 100 than at D = 30. The classical suite is tested at D = 30 and D = 50 only.