New hyper-heuristic algorithm automatically tunes its own learning period

In a new paper (arXiv:2605.29916), Benjamin Doerr, Pietro S. Oliveto, and John Alasdair Warwicker tackle a long-standing pain point in hyper-heuristics: the need to manually set a learning period parameter τ. Their Random Gradient hyper-heuristic automatically adapts τ on the fly, freeing users from guesswork. The authors prove that the resulting system selects the optimal neighborhood size for the LeadingOnes benchmark in a 1−o(1) fraction of all iterations. This leads to optimization time that matches the theoretical best possible (up to lower-order terms) for the neighborhood sizes used.

By removing human intervention from parameter control, the research bridges a key gap between theory and practice in evolutionary computation. The work relies on rigorous proofs rather than empirical tuning, offering a principled way to deploy meta-heuristics like Randomised Local Search (RLS) without sacrificing performance. For professionals dealing with pseudo-Boolean optimization (common in scheduling, logistics, and ML), this means faster convergence and less time spent on hyperparameter sweeps. The paper is set to appear in the journal Artificial Intelligence and has immediate implications for automated algorithm configuration.