Regularized Evolution
Algorithms AIMech:RegularizedEvolution DRAFT
Aliases: Aging Evolution
An evolutionary neural-architecture search algorithm that selects a parent by recorded validation accuracy from a uniformly sampled tournament, trains and evaluates a mutated child architecture, and replaces the oldest individual in its population rather than selecting the removed individual by fitness.
Identity Notes & Concept Scope
Current creation-context inventory at merged base fc27ccaf4668f1bf7468a10c269f5583fa3e640b covered 742 maintained and 442 pinned AIO records, including IDs, names, aliases and scoped synonyms, with no normalized candidate-probe collision. Four fixed-string word searches included ignored and hidden files and yielded 369/373/1045/1105 occurrences. Root read all eleven distinct matching strings and their selected provenance, while exact reconstruction accounted mechanically for all 2892 occurrences. Hits identify existing frontier leads and historical Fast AutoAugment code inventory paths, not a maintained record for this identity. The newly surfaced shakedrop.py path was listed at repository commit 2424224291616a2158a2265069ee95a78a2d13d9; no implementation was opened or executed. Prior primary-source, selected-neighbor and exact pinned CLASS/incident-edge inspections remain bounded by their individual research acceptances. This is not a global absence or saturation claim. This record denotes the named aging evolutionary architecture-search procedure: fitness tournament selection with replacement, architecture mutation and fresh training, followed by oldest-individual removal. Aging Evolution is its supported alias. Generic genetic optimization, weight-inheriting Population Based Training and augmentation-policy AutoAugment are neighboring but different scopes. The related TrainingStrategy association does not assert equivalence or a subclass relationship.
Claims & Grounded Evidence
6 Claims
Mechanism
#trained-population-and-architecture-search
The original procedure initializes a population queue with P model instances by generating random architectures, training each one and measuring its validation accuracy. Each instance retains its architecture and recorded score. In the studied NASNet search space, an architecture specifies the normal and reduction cells; repeated cells share the chosen cell structure. The procedure searches architectural choices rather than treating AmoebaNet-A, one architecture selected from its results, as the algorithm itself.
Mechanism
#with-replacement-fitness-tournament
At each update, draw S current population members uniformly with replacement and select the member with the highest stored validation accuracy in that sample as the parent. Repeated draws of the same individual remain possible. Parent selection depends on the recorded score, not an age bonus. S controls selection pressure: with S=1 the parent is sampled uniformly, but the next architecture still comes from mutating that parent rather than from the independent random-architecture baseline.
Mechanism
#whole-population-oldest-removal
After a child has been trained and evaluated, append it to the population and remove the oldest individual from the entire population queue. Thus the sequential procedure restores population size P at each completed update, after a transient size P+1. This age-based removal differs from removing the lowest-scoring member of the tournament sample. Age tracks individual residence in the population, not the gene-age layers of the separately discussed ALPS method. Repeated instances or descendants of an architecture are not forced to disappear merely because an older individual is removed.
Mechanism
#architecture-mutation-and-fresh-evaluation
The NASNet application forms each child by one randomly chosen architecture mutation. A hidden-state mutation changes an input connection in a selected pairwise cell combination while preventing cycles; an operation mutation changes the selected operation. An identity mutation can leave the architecture unchanged. The child architecture is then trained and evaluated again, including after an identity mutation; the procedure passes architectures rather than defining the inherited-weight exploitation used by population based training. The reported P=100, S=25 and identity-mutation probability 0.05 are study settings, not universal constants of aging evolution.
Mechanism
#evaluation-history-and-stopping
Keep the evaluated individuals in a history archive and return the individual with the highest recorded accuracy in that archive. The returned individual need not remain in the final live population. In the literal sequential pseudocode, initialization inserts P evaluated individuals into history before the condition |history|<C is checked. For integers C>=P>0, this gives C total evaluated individuals and C-P subsequent mutation updates. The archived maximum of fixed recorded scores cannot decrease as history grows; this does not imply that live-population scores or expected retrained performance improve monotonically.
Scope
#training-noise-motivation-and-empirical-boundary
The authors motivate aging as a way to limit the persistence of a lucky noisy training evaluation and to require further training evaluations along architectural lineages. They explicitly present that explanation as a conjecture. Their supplemental sanity check uses noisy bit strings rather than neural networks: the score is the fraction of zero coordinates plus Gaussian noise. That bounded example is circumstantial evidence for the motivation, not a general proof that aging removes training-noise bias, preserves diversity, reaches a global optimum or outperforms other search methods.
⚠️ Curated Limitations & Negative Results
- This record follows the original architecture-search procedure and its NASNet application. The host search space, mutation operators, validation protocol, training effort and final enlargement of selected models are separate choices. AmoebaNet architectures, ALPS gene-age layers and later platform-aware or weight-sharing variants are not exact aliases of the algorithm.
- Algorithm 1 uses an empty-set glyph for the tournament sample, while the explicit prose requires sampling with replacement. A list or multiset interpretation preserves that rule; a literal distinct-member set changes it. Ordinary operation requires positive integer P and S. Score ties and invalid-parameter behavior are not fully specified.
- The comment describing C evolution cycles is imprecise relative to the history guard, which includes the P initialization evaluations. If C<P, the literal code still initializes P individuals before performing no mutation updates. These sequential counts do not certify a hard budget under concurrent in-flight training jobs.
- The paper describes asynchronous parallel use but does not fully specify simultaneous completion order, stale population reads or queue concurrency. Finite lifetimes measured in completed sequential updates are not fixed wall-clock lifetimes, and removing an individual does not extinguish its architecture or lineage.
- The qualitative discussion that an inaccurately retrained model is never selected and its architecture disappears is not a literal guarantee of the tournament rule. A low-scoring member can be drawn repeatedly, other instances can remain and later mutations can recreate an architecture. The noise explanation does not establish general diversity, convergence, optimality or statistical debiasing guarantees.
- At S=1, uniform parent choice still induces population-dependent mutation, so it is not generally independent random architecture search. Likewise S=P with replacement does not guarantee sampling every population member. Uniform architectural encodings should not silently be interpreted as uniform distinct graph-isomorphism classes or trained functions.
- The core source is the official ten-page AAAI 2019 paper; the separately pinned arXiv v7 is used for Supplement C only. The publisher visible publication date, its machine metadata dates and the earlier 2018 workshop author page are distinct. Supporting research read the full official native text and method/discussion visuals, with selective supplement reading; it did not reproduce benchmarks, inspect author code or verify every supplement result. No numerical performance or universal speed claim is curated.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:TrainingStrategy
TrainingStrategy
|
related_match | The original search procedure organizes repeated model training and validation to select architectures. This supports a relation to the pinned training-strategy scope without adopting its preprocessing wording as the algorithm definition. The outer search algorithm is neither equivalent to nor asserted to be a subclass of that AIO target, a trained network or a layer. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
real2019-aaai
|
Esteban Real, Alok Aggarwal, Yanping Huang, Quoc V. Le
|
primary paper
full_text
|
ID: doi:10.1609/aaai.v33i01.33014780
sha256:39fa5f1d74dd…
|
real2019-arxiv-v7
|
Esteban Real, Alok Aggarwal, Yanping Huang, Quoc V. Le
|
primary paper
full_text
|
ID: arXiv:1802.01548v7
sha256:bc74c52b4c13…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-15T08:18:23Z
Updated: 2026-10-15T08:18:23Z
Source File:
data/curated/RegularizedEvolution.yaml