Home / Curated / Algorithms / DEHB

DEHB

Algorithms AIMech:DEHB DRAFT

Aliases: Evolutionary Hyperband

A hyperparameter optimization algorithm that combines Hyperband resource schedules with differential-evolution populations at different resource levels, using promising lower-resource configurations to guide proposals evaluated and selected at higher resource levels.

Identity Notes & Concept Scope DEHB is the combined resource-allocation and differential-evolution method, distinct from generic genetic algorithms, differential evolution alone, random-proposal Hyperband, KDE-guided BOHB and ASHA’s asynchronous within-stage promotions. Budget-specific populations contain configuration vectors, not a population of trained checkpoints whose weights are copied between configurations. The proceedings paper and the versioned appendix are distinguished explicitly. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 262 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 9 lines in each checkout. These were prior candidate leads, source references and neighboring-method mentions, not existing records for these identities. Ten exact AIO CLASS nodes with synonyms and outgoing edges and eight neighboring native mechanisms were inspected. Absence is bounded to this snapshot and semantic neighborhood.

Claims & Grounded Evidence

1 Claim

Mechanism #method

Maintain a differential-evolution subpopulation for each Hyperband resource level, sized according to the maximum number of evaluations assigned to that level. Iterate through the prescribed halving brackets. For a bracket’s starting level use evolution within its own population; for higher levels form a parent pool from promising lower-level configurations and use it for difference-vector mutation, followed by crossover with a target in the current level’s population. Evaluate the proposed configuration at the current level and retain it only according to a same-level fitness comparison with the target. Supplement an undersized mutation pool from a global pool of configurations. The initialization cycle includes direct lower-to-higher promotions before the general evolutionary transfer pattern. For mixed search spaces, the appendix keeps continuous population coordinates in a unit hypercube and maps copies to numerical or binned ordinal/categorical settings for evaluation. Its parallel design keeps each halving stage synchronized while allowing brackets to overlap and applying completed differential-evolution updates immediately.

supports Source: DEHB: Evolutionary Hyberband for Scalable, Robust and Efficient Hyperparameter Optimization · Locator: Sections3–4.3, PDF pp.2–4 (printed2148–2150), especially Figures1–3 and the modified halving procedure
supports Source: DEHB: Evolutionary Hyperband for Scalable, Robust and Efficient Hyperparameter Optimization · Locator: AppendicesA–C, PDF pp.9–12, especially Algorithm2 on p.10, Algorithm3 and SectionsC.2–C.3 on pp.11–12; title/version on p.1

⚠️ Curated Limitations & Negative Results

The source’s initialization prose and pseudocode use bracket and iteration terminology inconsistently. Algorithm3’s bracket_counter==0 condition spans an outer cycle and applies direct promotion only for i>0. The record describes the common mechanism without treating every printed initialization statement as byte-equivalent executable steps.
Appendix Algorithm2 prints b0=(bmax/bmin)*eta^(-s), which gives normalized rather than absolute budgets when bmin differs from1. Its return order for counts and budgets is reversed at the Algorithm3 call site. These discrepancies are retained as source qualifications; no author-issued erratum or exact executable schedule is claimed.
The appendix updates its incumbent after evaluations without an explicit full-budget restriction. The record therefore does not promise that the returned best observed configuration has been verified at maximum resource. Greedy same-budget replacement limits harmful transfers but is not a proof of global optimality or universal robustness.
Continuous encodings impose numerical neighborhoods on categorical choices, and low-fidelity information can be misleading. Parallel completion order can affect shared populations. No benchmark speedup, constant total runtime or author-code execution is claimed.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:TrainingStrategy` TrainingStrategy	related_match	This hyperparameter procedure selects model configurations and allocates training resources. The relation links that scope to the pinned training-strategy concept without treating the optimizer as preprocessing, an identical concept or a subclass.

Evidence Sources & Primary Literature

2 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`method-source`	DEHB: Evolutionary Hyberband for Scalable, Robust and Efficient Hyperparameter Optimization Noor Awad, Neeratyoy Mallik, Frank Hutter	primary paper full_text	ID: doi:10.24963/ijcai.2021/296 sha256:48ff58a4bdc3…
`versioned-appendix`	DEHB: Evolutionary Hyperband for Scalable, Robust and Efficient Hyperparameter Optimization Noor Awad, Neeratyoy Mallik, Frank Hutter	primary paper full_text	ID: arXiv:2105.09821v2 sha256:90b2d08b7234…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T11:53:28Z

Updated: 2026-10-11T11:53:28Z

Source File: data/curated/DEHB.yaml