Skip to content
AMM

AMM (Adaptive Multi Mode) optimizer

Adaptive Multi Mode · derivative-free black-box optimization

AMM reads your function with cheap probes, then routes the evaluation budget to the right classical algorithm, so one setting works across smooth, rugged and multimodal problems.

pip install amm-optimizerPyPI: 0.3.0 on PyPI (0.2.0 yanked).
import numpy as np
import amm_optimizer


def rosenbrock(x):
    return float(np.sum(100 * (x[1:] - x[:-1] ** 2) ** 2 + (1 - x[:-1]) ** 2))


result = amm_optimizer.minimize(
    rosenbrock, x0=np.zeros(5), bounds=(-5, 5), budget=2000, seed=0,
    config=amm_optimizer.presets.get("c46"),
)
print(result.nfev, result.fun)
print(result.x.round(3))
2000 2.4585372370019567e-13
[1. 1. 1. 1. 1.]

The known optimum is x = (1, 1, 1, 1, 1) with f = 0. config= selects AMM (the frozen configuration c46); without it minimize runs the original ATM sampler, which reaches only f = 1.7 on this example. That output was produced by the frozen configuration c46 (config hash 2153948ebed51150) on 2026-10-22. Source and issues: https://github.com/BjBodner/amm-optimizer.

How it compares

Section titled “How it compares”

On bbob instances that no design decision used (ids 101-110), AMM has the highest mean share of targets reached in 382 of 600 function, dimension and budget cells (64%, ties included). It is credibly behind some competitor in 35 (5.8%), and all 35 are listed on the limits page. It leads pycma IPOP-CMA-ES by +.12 to +.31 at every budget, and restarted SLSQP at every budget too, by +.015 [+.007, +.022] at 10 times the dimension, growing to +.302 [+.246, +.362] at 1000 times the dimension. The field is pycma IPOP and BIPOP, SLSQP, differential evolution, Optuna’s TPE, Nevergrad’s NGOpt and HCMA. The measure is the share of targets reached, 0 to 1.

382of 600 cells: AMM has the highest mean
35of 600 cells: credibly behind some competitor
110of 600 cells: credibly ahead of every competitor that ran it

This is the preregistered final test. The frozen configuration was run once on these instances. The analysis was run twice, as logged in advance: a first version without NGOpt at D 5 to 40 (still running then), and this one with every competitor at every dimension. “Credibly” is the rule fixed in advance: a competitor leads by more than 0.02 with posterior probability above 0.9 over the ten paired runs. The intervals behind it are about 25% too narrow, so these are counts and not significance statements. Every preregistered prediction for the final test held; the research log has them and the history of the analysis.

AMM (c46)pycma IPOPpycma BIPOPSLSQPDETPE (Optuna)NGOpt (Nevergrad)HCMA (COCO archive)
Five line charts, one per dimension (2, 5, 10, 20 and 40), of the fraction of function-target pairs reached against budget from 10 to 1000 times the dimension, on the final-test instances. AMM's thick line is the highest, or level with the highest, at every budget in every chart. At the largest budget the nearest competitors are differential evolution, HCMA's archive levels and the CMA-ES variants in two dimensions, HCMA's archive levels in 5 to 20 dimensions, and restarted SLSQP in 40 dimensions. NGOpt, drawn at every dimension, is below AMM everywhere; in 5 dimensions it reaches about half of the pairs at 300 times the dimension and falls back at 1000.
final testFinal test FT1: fraction of (function, target) pairs reached versus budget, bbob f1 to f24, one panel per dimension, never pooled across dimension. Grey band: the best method in each cell. HCMA is shown at the COCO archive's own levels (D 2 to 20, other instances), not rerun. NGOpt is shown at every dimension.
Reproduce this figure
command
final_ft1.py with no arguments, and ft1_plots.py as run for v2 (batch b095_final_ft1)
seed
9000, one run per function, dimension, instance and budget (6,000 runs of AMM)
instance ids
101-110, never used by any design decision
budgets
10, 30, 100, 300, 1000 x D
commit
frozen configuration c46, config hash 2153948ebed51150 (the tagged history is in the research archive)
competitors
pycma IPOP/BIPOP, SLSQP, DE, TPE (capped slice) and NGOpt in the same harness; HCMA from the COCO archive
hardware
pending: filled from the run manifest
cost
pending: filled from the run manifest
version
0.3.0
config hash
2153948ebed51150
licence
Apache-2.0
repository
https://github.com/BjBodner/amm-optimizer
PyPI
0.3.0 on PyPI (0.2.0 yanked)
research archive DOI
pendingpending
checks
local gate tools/ci.py, 100% coverage, golden-master tests

What is in here

Section titled “What is in here”