Changelog

The public repository (https://github.com/BjBodner/amm-optimizer, commit c4f0ab1) is version 0.3.0. PyPI: 0.3.0 on PyPI (0.2.0 yanked). Its own CHANGELOG.md has two entries, 0.3.0 and 0.2.0, which are condensed first. The development history below them is condensed from the development changelog, which is part of the tagged history in the research archive (DOI: pendingpending, pending), not of the public repository.

0.3.0 (2026-10-23)

A packaging release; no algorithmic change. The configuration c46 (config hash 2153948ebed51150) is bit-identical to the frozen one.

• Breaking: the import name is now amm_optimizer. 0.2.0 installed the top-level modules atm and amm, which collide with unrelated PyPI projects of those names; both modules are gone. ATMConfig and the preset names (c46) are unchanged.
• The default configuration is unchanged. amm_optimizer.minimize without config= runs the original ATM sampler (not AMM). AMM needs config=amm_optimizer.presets.get("c46").
• The package requires Python 3.12 or newer, NumPy 2.0 or newer and SciPy 1.18 or newer. 0.2.0 declared lower floors that SciPy 1.16 or newer cannot satisfy.
• Fixed: import no longer fails on NumPy 1.x; the c46 preset no longer crashes for one-dimensional problems at budgets of 150 or more.
• Added: py.typed, an installed-package smoke test, GitHub Actions workflows (CI, release, security checks), SECURITY.md and CONTRIBUTING.md.

0.2.0 (release of the freeze)

Packaging of the frozen code: import as amm (the implementation kept its code name atm), CITATION.cff and an AMM README. No algorithmic change. It was published to PyPI and has since been yanked: its default minimize is the original ATM sampler and it imports as amm or atm. 0.3.0 replaces it.

Development history before 0.2.0

The entries below describe the code as it was at the time, when the module was atm.

2026-10-22: code freeze

• The configuration c46 is frozen for the paper (docs/FREEZE_2026-10-22.md, config hash 2153948ebed51150). The algorithm does not change for the final test.
• Cycle 71: an opt-in option stage_extend_max_per_dim (the stage_extend_restart option only up to a budget per dimension). The capped version was not confirmed on fresh ids, so it is not in any preset.
• Batch 91 was rejected under the letter of its pre-set rule. c46 is unchanged.

2026-10-22: cycle 70, c46

• New preset c46 (c45 plus prefix_low_beta=0.15 from dimension 10): in the evolution-strategy prefix at 100 x D, a finite rugged curvature reading below the threshold keeps the (1+1)-CMA-ES instead of the lq2 populations (aimed at bbob f23). Confirmed on fresh ids and promoted. It is the best composite.
• Opt-in options added and not in a preset: centre_point / centre_point_min_dim, smooth_bfgs_second.
• Batch 89 (stage_extend_restart on c45) was rejected.

2026-10-22: cycles 67-69, c45

• New preset c45. It adds a rugged-route BFGS try that continues to the end of the budget on a kink reading, a ridge reading that hands the rest of the prefix to restarted COBYQA, retries of a descent that did not move, a band where the finisher’s populations start at once, and box_normalize. Confirmed on fresh ids.
• Variant.box_normalize: minimize maps any box onto [-5, 5] in each coordinate and runs there. It is bit-identical on bbob, where the box is already [-5, 5].
• Several other options were added as opt-in and left out of the presets, or rejected, among them sampled_polish_per_dim and sep_route_low.

2026-10-21: cycles 63-66, c44

• New preset c44: plateau populations in the prefix from dimension 5, and small_switch, which reads the first SLSQP descent at two checkpoints at budgets up to 10 x D and may hand the rest to restarted COBYQA.
• Variant.prefix_nan_reading: where an undefined curvature reading of the prefix goes. Opt-in.

2026-10-21: cycles 52-61, presets c39 to c43

• c43: atm.cobyqa_phase.run_cobyqa_restarts (restarted SciPy COBYQA in the box) and the small-engine option for the smallest budgets, in dimensions up to 3.
• c42: after basin_route’s smooth, single-basin reading at dimensions 2 and 3, SLSQP descents from the design, then the usual path.
• c41: qn_phase.run_scaled_bfgs(box_qp=...), SLSQP’s bound handling in a BFGS descent, and the route qn44.
• c40: a one-descent prefix (one SLSQP descent, then the (1+1)-ES) at small budgets.
• c39: atm.basin_phase: a Sobol design and a curvature probe send smooth multi-basin readings (Gallagher-type functions) at dimensions 2 and 3 to MLSL; a stall check on the separable search.
• Variant.fork_seed (research): independent continuations from one hand-over, for a router dataset.

2026-10-20: cycles 35-51, presets c35 to c38

• c38: atm.chain_phase and the options chain_route and sep_route. A pair test of additivity sends a linear chain (bbob f20) to a decoupled coordinate search, and a separable function to restarted STEP.
• c37: a restarted nonsmooth BFGS on forward differences (qn_phase.run_scaled_bfgs) for the f13 kink route and a capped try at dimensions 5 to 9.
• c36: a fractal-reading route to restarted SLSQP, and prefix switches (separable smooth prefix reaches SLSQP).
• c35: Gallagher-style restarts of SLSQP from the best unclaimed sample point.
• step_phase.run_step_phase keeps its samples in arrays: bit-exact and much faster at large budgets.

2026-10-19: integration of the research line, presets c23 to c34

• The research line (composites c23 to c31), the code-review branch and refactor were merged. Every characterised run of the research tip reproduces bit for bit, except 14 round-1 arms that use a removed option (docs/MERGE_2026-10-19.md).
• atm.presets holds the named configurations, with get, names and with_changes. Each preset equals benchmarks/arms/<name>_config.txt, and quick_scan --config preset:NAME reads them.
• Renames: Variant.nm_max_dim of refactor became simplex_max_dim; the research line’s ATMConfig.improved() became ATMConfig.improved_wave5(), and improved() is the Tier 5 hybrid.
• c32 to c34 added the L-BFGS-B fallback when an SLSQP call fails more often than it converges, an IPOP route from dimension 20, a stall monitor on the first SLSQP descent, and a second descent from a uniform point.
• Harness fix: the cache of bbob optimal values (f_opt) held wrong values for 291 f12 instances because of rounding in a COCO data line. The harness now reads the exact value, repaired the cache, and flags runs that stopped on a false final-target hit. Every f12 score before this date is affected.
• scipy>=1.16 became a declared dependency.
• Research tooling added: a posterior goal ledger, fresh-id claim scripts and a consolidated PDF builder.

Before the research line (undated entries)

The repository changelog does not date these entries.

• ATMConfig.variant (atm.config.Variant): opt-in algorithm variants, every default equal to the original and tested bit for bit; ATMConfig.to_dict, from_dict, config_hash; minimize(bounds=...).
• ATMConfig.improved(): the best configuration of docs/IMPROVEMENTS.md (needs bounds=); ATMConfig() stays the original. Later waves added a (1+1)-ES prefix, a learned covariance, a late Nelder-Mead refinement, and a wave-7 restriction of the local stages after a holdout found a regression on multimodal functions. improved_wave6() keeps the previous preset for reproducibility.
• A scoring fix (exact bbob optima instead of cached values that floored converged runs), raw COCO evaluation and the 2019 start protocol; benchmarks/quick_scan.py version 2 (stable seeds, provenance, process pool, resume, cluster bootstrap).
• The COCO/bbob harness (benchmarks/run_bbob.py and companions, docs/BENCHMARKS.md), block profiling (atm.profiling), a CI workflow, a py.typed marker, and faster parent selection (bit-identical).
• Opt-in variants that were added and then removed again for no aggregate gain are recorded in docs/DEVIATIONS.md.

0.2.0.dev0 (development version)

• A clean re-implementation of ATM as the atm package, bit-identical to BjBodner/atm@69828c6 with ATMConfig(strict_budget=False).
• Strict budget by default (at most budget calls, all counted), handling of NaN and infinite objective values, conductor robustness, and an opt-in correct second best. All differences from the original are listed in docs/DEVIATIONS.md.