ARROWDE: an adaptive rotation-safe routed operator workflow for differential evolution

Optimizing continuous problems that combine rotated, non-separable, and multimodal landscapes under a limited budget of function evaluations is difficult. Most Differential Evolution (DE) methods use a static strategy or a fixed operator mixture, which can converge prematurely or spend the budget poorly when problem properties change during a run. This study proposes Adaptive Rotation-Safe Routed Operator Workflow Differential Evolution (ARROWDE). This bandit model combines upper-confidence-bound learning with landscape-aware priors, routing each trial to one of six single-evaluation operators: three exploratory and three exploitative. Implicit orthogonal transformations built from Givens rotations give robustness on rotated problems, and rotation-invariant population statistics are cached for the direct search operators at low cost. . On the CEC 2014 and CEC 2017 suites at 30 and 50 variables, ARROWDE attains the best mean Friedman ranks among the compared methods and is favored by the suite-level comparison against the most recent competition winner, with the margin widest at the higher dimension. The same tables mark its boundaries: IMODE and LSHADE-EpSin achieve lower mean errors on the 30-variable CEC 2014 composition functions, and the framework runs more slowly than L-SHADE and CMA-ES. Rather than proving every component strictly indispensable, ablation study ranks their relative importance, with stripping the rotation-safe layer causing the steepest performance drop, followed by the landscape prior and inherited adaptation, whereas the caps and bandit rule exhibit the smallest overall impact.