ALEX-GWO: Reinforcement-learning-driven Adaptive Grey Wolf Optimization with Landscape-aware Local-optima Escape for Large-scale Global Optimization
Abstract
Large-scale global optimization exposes a structural weakness of the grey wolf optimizer: its convergence coefficient follows a schedule fixed in advance, the three leading wolves collapse into a single attraction basin long before the budget is spent, and that loss of diversity is neither detected nor repaired.Published variants answer the symptoms with fixed remedies applied on a schedule rather than on evidence, so their cost is paid whether or not the population is trapped.This paper presents ALEX-GWO (Adaptive Learning-driven EXploration Grey Wolf Optimizer), which closes the loop between observation and control.A six-component search-state vector summarises progress, diversity, success rate, improvement, stagnation and leader dispersion; population-level parameter control, per-wolf success memory and an epsilon-greedy Q-learning selector consume it at three time scales; and a composite landscapetrapping score releases a targeted escape only when several independent indicators agree.On CEC'2013 and CEC'2010 at 1000 dimensions under an equal budget of 3.0 x 10^6 evaluations, ALEX-GWO obtains the best average Friedman rank of the ten algorithms compared on the first suite (2.00) and of the nine on the second (1.85)a set including ADSL-GWO, RL-LGWO and IAGWO re-implemented under an identical protocoland improves on canonical GWO by a median of four orders of magnitude.It reaches a lower mean error than SHADE-ILS on 13 of the 15 CEC'2013 functions (Wilcoxon p = 0.048) but is not separated from MOS, CC-RDG3 or IAGWO after Holm correction, so the evidence supports competitive consistency rather than dominance.ALEX-GWO costs 11.3% more runtime than canonical GWO, and removing landscape detection alone drops the escape success rate from 0.68 to 0.31.A binary variant reaches 90.32-93.75%accuracy under nested cross-validation on four feature-selection datasets while discarding 65-97% of the features.Code and raw results: https://github.com/Rajpreet-kaur1989/ALEXGWO.
