Home / Curated / Algorithms / BI-population CMA-ES

BI-population CMA-ES

Algorithms AIMech:BIPOPCMAES DRAFT

The original 2009 multistart CMA-ES algorithm that interlaces increasing-population attempts with randomized smaller population and initial-scale attempts according to spent function-evaluation budgets, using its specified positive-weight covariance and cumulative step-size base.

Identity Notes & Concept Scope This record describes the original 2009 positive-weight whole BIPOP method with its specified CMA base, spent-evaluation-budget interleaving and independent random small-population/initial-scale rules. It is distinct from the existing 2005 IPOP algorithm, 2001 rank-one base, 2003 covariance component and 2006 signed active component. Fresh postmerge collection at 1031c6b1f1f4492fbebec6eee1db5a6c68672d9e checked all 737 native and 442 pinned AIO record identities and bytes. Four searches included ignored and hidden files, yielding 1,463 occurrences grouped into 26 exact contents; root read all 26 complete contents. No normalized ID, label, alias or scoped-synonym collision was found for the explicit probes. The matches are carried active-covariance leads, unrelated Gaussian-process cross-covariance language, and existing IPOP boundary context. This finding is bounded to the pinned inventory and inspected semantic neighborhoods, not all AI literature. Existing owner 202 retains its IPOP record and prior resolution, adds BIPOP, and keeps other restart residuals.

Claims & Grounded Evidence

6 Claims

Mechanism #fixed-positive-weight-cma-base

The original whole algorithm uses the specified Gaussian CMA-ES base for minimizing an evaluated objective in integer dimension D>=1. Each generation samples lambda independent offspring with mean m and covariance sigma^2*C, ranks them by objective value, and recombines the best mu=floor(lambda/2) with positive normalized weights. The author erratum specifies the weights actually used as w_i=(ln(mu+1/2)-ln(i))/sum_j(ln(mu+1/2)-ln(j)), with i,j from 1 to mu. The covariance rule combines retained C, a rank-one evolution-path term and a weighted second moment of selected displacements about the old mean. Its h_sigma gate suppresses the new covariance-path increment when the step-size path is too long and adds the stated correction to the old-C coefficient. Population-dependent weights, effective selection mass and learning rates are recomputed for the population used by each attempt.

supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed · Locator: Section 2, PDF p. 1 / printed p. 2389, constants and recombination weights; Table 1, PDF p. 2 / printed p. 2390, caption and Equations 1–5.
supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed — author erratum · Locator: Minor erratum: the computation of w_i in Section 2 used ln(mu+1/2) rather than the printed ln(mu+1).

Mechanism #inherited-cumulative-step-size-adaptation

The fixed base accumulates the normalized mean displacement after whitening by the symmetric inverse square root of the current positive-definite C into a separate path p_sigma. It updates the positive scalar scale by sigma_new=sigma*exp((c_sigma/d_sigma)*(norm(p_sigma_new)/E_norm_standard_normal-1)), where the expectation is the length of a D-dimensional standard-normal vector. This inherited scalar exponential is distinct from the restart budget policy and population doubling. In exact real arithmetic, finite exponents preserve a positive sigma; this does not establish objective improvement or finite-precision stability.

supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed · Locator: Section 2, PDF p. 1 / printed p. 2389, inverse square root and step-size constants; Table 1, PDF p. 2 / printed p. 2390, Equations 3 and 6.

Mechanism #spent-budget-restart-selection

After one standalone run with lambda_def=4+floor(3 ln D), the procedure interlaces complete attempts from two restart regimes. Each regime has an account of function evaluations already spent; the initial standalone run is disregarded in the large-regime account. A small-regime attempt is selected if and only if its current spent budget is smaller than the large-regime budget; otherwise the large regime is selected, including ties. The first restart is therefore large. This is evaluation-cost balancing between sequential attempts, not simultaneous evolution of two populations, fixed alternation, equal numbers of attempts or exact equality of costs after every attempt.

supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed · Locator: Section 2.1, PDF p. 1 / printed p. 2389, initial run, budget accounting, first restart and strict small-budget selection condition.

Mechanism #increasing-population-regime

In the original large-population regime, lambda doubles before each large restart, with initial scale sigma_0=2. The configured sequence permits at most nine such restarts, so its largest population is 2^9*lambda_def. These are large-regime restarts, not a bound of nine on all attempts including the interlaced small regime. Doubling occurs between attempts rather than every generation. The source describes a schedule whose final configured restart is large; stopping early on the objective target can instead finish in a small attempt. The nine-restart setting is specific to this original configuration.

supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed · Locator: Section 2.1, PDF p. 1 / printed p. 2389, increasing population and first/last restart statements; Section 2.2, PDF p. 2 / printed p. 2390, overall target or largest population stopping; Section 6 and Table 3, PDF p. 3 / printed p. 2391, final successful populations.

Mechanism #randomized-small-population-and-scale

Given the latest large-regime population lambda_l, the small regime uses independent uniform draws U,V on [0,1] to set lambda_s=floor(lambda_def*(lambda_l/(2*lambda_def))^(U^2)) and sigma_0=2*10^(-2V). Thus lambda_s lies between lambda_def and lambda_l/2, while the initial scale is log-uniform on [0.02,2]. At the first large population 2*lambda_def, the small population is simply lambda_def; later large populations permit variation. The population rule is not a uniform draw over integers in that range. The paper also enforces a small-run evaluation cap described as half the recent large budget; this record preserves that wording without certifying a last-run versus cumulative implementation convention.

supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed · Locator: Section 2.1, PDF p. 1 / printed p. 2389, displayed lambda_s formula, independent-uniform statement, sigma_0 formula and recent-large-budget cap. Range and first-large-population observations follow directly from these formulas.

Scope #original-configuration-and-stopping-boundary

In the evaluated configuration, each run starts with a mean sampled uniformly in [-4,4]^D, identity covariance and zero evolution paths. That box specifies initialization, not a claim that every Gaussian offspring remains inside it. A local attempt ends on its target or the listed iteration, function-range, stagnation, small-coordinate/path, no-effect or covariance-conditioning criteria. Overall termination also depends on the target or largest configured final population. These controls are not global-optimality certificates. This named positive-weight 2009 whole method is distinct from the original increasing-population-only restart procedure, a covariance-update component or a generic controller with an arbitrary inner optimizer.

supports Source: Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed · Locator: Table 1 caption and Section 2.2, PDF p. 2 / printed p. 2390, initialization and termination; Section 2.1, PDF p. 1 / printed p. 2389, the two different regimes; Section 3, PDF p. 2, configuration qualifications.

⚠️ Curated Limitations & Negative Results

This is the original 2009 positive-weight whole method, with the separately cited correction to its recombination weights. The maintained 2001 rank-one base, original 2003 equal-weight rank-mu component, 2006 active signed-moment component and 2005 increasing-population-only procedure are different scopes. Later active-BIPOP combinations, noisy-objective extensions, alternative restart controllers and current software are not exact aliases of this record.
The paper defines cumulative spent-evaluation accounts but does not fully settle the executable meaning of half the recent large budget, partial-generation cap enforcement or budget overshoot. Equal-budget language does not guarantee finite exact equality. The configured first/last-large schedule is qualified by possible early target stopping.
Section 2.2 prints the small-path test without explicit absolute-value syntax and does not completely specify fitness-tie handling. No executable repair is asserted. The initialization box, scale choices, finite thresholds and known-target benchmark conditions are configuration choices, not assumptions satisfied by every black-box problem or proofs of invariance of the whole stopped run.
Table 1 gives the update equations, while Section 4, Equation 7 on PDF p. 3, describes delayed eigendecompositions and slightly outdated factors in the evaluated implementation. Exact refreshed factors at every generation, numerical conditioning, stable floating-point behavior and complete implementation fidelity are not certified here.
Section 3 reports no thorough parameter study and describes the small-regime settings as ad hoc. The reported evaluation totals and selected population/scale outcomes do not establish a universal hard evaluation cap, optimal budget allocation, global convergence or superiority over every restart policy. No performance number, speedup or author-code benchmark reproduction is curated.
The author erratum states that Table 2 RT_succ averages Nsucc arbitrary trials rather than necessarily successful trials, and that its final unsuccessful-trial median entry can be up to 12 percent too small. These corrections are retained without inventing repaired results or treating them as a failure of every method equation.
The primary bytes are a seven-page university-hosted proceedings copy, with printed pages 2389–2395; byte identity to a publisher or HAL copy is not certified. The author bibliography and erratum agree with those pages, while an archive listing gives 2389–2396. File build dates are distinct from publication metadata. Supporting research read all seven native-text pages and method/discussion visuals on pages 1–3, plus the complete author erratum. Empirical plots, unread bibliographic dependencies and the separate noisy companion were not audited. No author implementation, calculator, optimizer or benchmark was run.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ExponentialFunction` Exponential Function	related_match	The specified base of the original whole algorithm requires the scalar exponential in cumulative step-size adaptation, Table 1 Equation 6. This inherited component supports the related mapping; budget balancing and population doubling are not its basis. The optimizer is neither equivalent to nor a subclass of the scalar function, and no neural activation or elementwise matrix-exponential interpretation is asserted.

Evidence Sources & Primary Literature

2 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`hansen2009bipop`	Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed Nikolaus Hansen	primary paper full_text	ID: https://doi.org/10.1145/1570256.1570333 sha256:bdeda80ecf49…
`bipop-author-erratum`	Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed — author erratum Nikolaus Hansen	official documentation documentation	sha256:d41ce85e51d0…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-15T06:52:10Z

Updated: 2026-10-15T06:52:10Z

Source File: data/curated/BIPOPCMAES.yaml