Adaptive Operator Portfolio Search with Fold Based Landscape Probing

Abstract

Adaptive operator selection allocates search effort among candidate mechanisms whose usefulness changes with the current population. We study AOP-KFP, a portfolio of nine operators with periodic probing of held-out parents, exponentially smoothed improvement credits, and a softmax distribution mixed with uniform exploration. The folds organize local comparisons within a population snapshot; they do not establish statistical generalization to unseen problems. We give an explicit accounting of probe evaluations and a conditional hitting-probability result for elitist search, while distinguishing these statements from finite-budget performance. The supplied CEC 2017 table contains 25 functions at dimension 30 with 51 reported runs per algorithm. Its displayed means give AOP-KFP 17 row minima, jSO three, EBOwithCMA four, and L-SHADE one. Additional supplied rank, significance, ablation and sensitivity summaries require reconciliation with their underlying observations: several corrected-p-value entries are arithmetically inconsistent, and the reported Sobol first-order indices sum to more than one. The study therefore identifies a testable operator-allocation mechanism and descriptive benchmark evidence, with evaluation accounting and raw-output reconciliation required before confirmatory superiority or reproducibility claims can be made.