Paper and cite

Adaptive Multi Mode: Budget-Aware Routing of Optimizers By Online Landscape Probes

Status	In preparation. Not yet posted.
arXiv	pendingpending
PDF	pendingpending. The link appears when the preprint is posted.
Authors	Benjamin Jacob Bodner
Code	https://github.com/BjBodner/amm-optimizer (public, version 0.3.0, commit c4f0ab1)
Research archive	Preregistrations, FT1 data and outputs, the id ledger and the tagged history. DOI: pendingpending

Some results are still pending

The abstract below is adapted from the current draft. The final test (FT1, ids 101-110) is complete, with NGOpt at all dimensions (v2), and every preregistered prediction of it held. Results beyond bbob are not in yet. Every sentence that depends on those carries a marker and is not final. This page cannot be published while any marker remains.

Abstract

The best classical optimizer for a continuous black-box problem depends on the budget and on the landscape, but existing routers decide from the dimension and budget alone, from which engine improved last, or from features fitted to one benchmark. AMM routes a known budget among quasi-Newton, trust-region, coordinate, multistart and evolution-strategy engines with online probes of curvature, plateaus, separability, kinks, basins and fractal structure. Each probe has an engine mapping, a default outcome and a dimension-by-budget validity range, and its evaluations count against the budget.

AMM was built on bbob in a loop that diagnoses a losing cell, preregisters a fix and confirms it on never-used instances (24 of 27 predictions held). One configuration was then frozen and tested once on the official instances 101-110 and beyond bbob (not yet testedpending). During development, an intermediate configuration led IPOP-CMA-ES on never-used instances by +.208 [+.195, +.222] in target share. In the final test, AMM’s mean target share exceeds IPOP-CMA-ES’s by +.12 to +.31 at every budget (budgets of 10 to 1000 evaluations per dimension; Holm-corrected, bootstrap intervals above zero). AMM has the highest mean target share, ties included, in 382 of 600 function, dimension and budget cells (64%). It is credibly behind some competitor in 35 cells, all listed, and credibly ahead of every competitor that ran the cell in 110. Against HCMA’s COCO archive levels (D 2 to 20, other instances) AMM is ahead from 30 evaluations per dimension: +.120 at 30 x D, +.174 at 100, +.104 at 300 and +.103 at 1000 x D; the margin at 10 x D is small (+.016). AMM is an open-source Python package, installable with pip install amm-optimizer (PyPI release: 0.3.0 on PyPI (0.2.0 yanked)).

Adapted from the paper draft of 2026-10-21. The final-test result is stated as the number of cells where AMM has the highest mean target share (ties included), together with the preregistered count of credibly behind cells. The abstract on this page was updated to the final-test numbers of this site.

Contributions

• AMM, a budget-aware router among classical algorithms driven by online probes with stated costs, defaults and validity ranges, each traced to the function that motivated it.
• Held-out evidence for one frozen configuration on a ladder of controls, from each engine alone to the per-cell oracle, with its losses named. Results (final test): the highest mean in 382 of 600 cells and credibly behind in 35; the limits page lists those cells.
• The development loop as a method in its own right, with its record: 24 of 27 preregistered predictions held in the last twelve confirmations (see the research log).
• An open-source Python package, installable with pip, and COCO data for every budget multiplier. Release: 0.3.0 on PyPI (0.2.0 yanked).

Where to find things

In the paper	On this site
Method: probes, engines, routing	How it works
Results and the field comparison	Results (the final test and the in-sample development context)
Where AMM does not win; limitations	Limits and honesty
Development process, preregistrations, the final test FT1	Research log
Software availability	Get started, Reproduce

Cite

There is nothing to cite yet. A preprint identifier and a DOI will be set when they exist. The entries below are templates with the pending fields marked PENDING; do not copy them into a manuscript before then.

Pending now: arXiv identifier pendingpending, DOI pendingpending, year pendingpending.

BibTeX

@misc{amm_optimizer,
  title         = {{Adaptive Multi Mode: Budget-Aware Routing of Optimizers By Online Landscape Probes}},
  author        = {Bodner, Benjamin Jacob},
  year          = {PENDING},
  eprint        = {PENDING},
  archivePrefix = {arXiv},
  doi           = {PENDING},
  note          = {Preprint}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use AMM, please cite the paper below."
title: "AMM (Adaptive Multi Mode) optimizer"
version: 0.3.0
license: Apache-2.0
repository-code: "https://github.com/BjBodner/amm-optimizer"
authors:
  - family-names: Bodner
    given-names: Benjamin Jacob
preferred-citation:
  type: article
  title: "Adaptive Multi Mode: Budget-Aware Routing of Optimizers By Online Landscape Probes"
  authors:
    - family-names: Bodner
      given-names: Benjamin Jacob
  year: PENDING
  doi: "PENDING"

The repository has its own CITATION.cff for the software (version 0.3.0); the paper reference will be added when it is public. The software citation will use the DOI of the first release (pendingpending). The research archive has its own DOI (pendingpending).