HAMO

Mechanism

A computer-implemented meta-optimizer for black-box continuous minimization at high problem dimension. An O(D) probing protocol spends at most about six percent of the total function-evaluation budget characterizing the objective landscape: Latin-hypercube sampling, centroid-anchored random-direction line segments, a quadratic radial fit, and short multi-start quasi-Newton descents. A trained classifier reads the probe features and assigns a landscape class.

The remaining budget is routed to a component optimizer selected from a fixed pool: separable CMA-ES, BIPOP-CMA-ES, restarted L-BFGS-B, and a graduated-smoothing zeroth-order method derived from the CMLGS mechanism. The routed component executes from a neutral centroid start under a no-warm-start rule, a bounded fallback restart phase guards the result, and a quasi-Newton polish phase sharpens the smoothed-landscape minimum. The method is operable at dimensions 80 to 640 on the BBOB-largescale 2019 suite.

Prior-art differentiator

Classifier-routed algorithm selection is established in the literature, but at low dimension on the standard BBOB testbed. The dominant exploratory-landscape-analysis feature sets scale quadratically in problem dimension and become empirically infeasible at the dimensions where this method operates. The O(D) probe feature set is the engineering discovery that closes that gap: landscape characterization cheap enough to run at dimension 640 inside a single-digit-percent budget envelope.

Separately, no verified portfolio-routing reference includes a graduated-smoothing zeroth-order method as a routed component. The two hooks jointly distinguish the claim: high-dimensional operability and the graduated-smoothing component inside the routed pool. Probe-budget-fraction analyses exist in recent prior art at far lower dimension and without the graduated-smoothing component; learned-schedule and slice-probing portfolio systems are structurally distinct from fixed-pool classifier routing.

Validation scope

Across the BBOB-largescale 2019 benchmark at dimensions 80 to 640, the routed configuration records the highest top-2 frequency and the lowest mean rank in a six-algorithm comparator field. The as-filed specification discloses the full pre-registered success-criteria record, including where strict thresholds were not met.

Counsel posture

Provisional self-filed 2026-05-27 through USPTO Patent Center. The non-provisional conversion window runs to 2027-05-27. Assignment recordation under 37 CFR 3.11 is in progress.