Natural Evolution Strategies
Algorithms AIMech:NaturalEvolutionStrategies DRAFT
Aliases: NES
A family of black-box optimization algorithms that adapts a parameterized search distribution using estimates of the natural gradient of expected fitness with respect to the distribution parameters.
Identity Notes & Concept Scope
NES is the shared natural-gradient search-distribution framework, not a single Gaussian implementation, a trained neural network or ordinary stochastic gradient descent on model weights. Its canonical method differentiates the search distribution’s log density, not the black-box objective itself. Exponential NES and separable NES are concrete variants with additional parameterization and computational choices; the framework record does not equate them. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 270 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in worktree and main, returning 19 and 19 lines respectively. These were prior candidate leads, source discussions and neighboring DEHB/FABOLAS method mentions. Eight exact AIO CLASS nodes with scoped synonyms and outgoing edges and seven native neighboring mechanisms were inspected. Absence is bounded to this snapshot and semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Draw candidate points from a parameterized search distribution and evaluate their fitness. Compute the log-density derivative, or score, of each sample with respect to distribution parameters. In the canonical raw-fitness formulation, estimate the expected-fitness gradient by averaging score vectors weighted by fitness. Obtain the Fisher information of the distribution, analytically when available or through score outer products, and use its inverse action to transform the search gradient into a natural-gradient direction. Update the distribution parameters with a chosen learning rate and repeat. The paper additionally describes rank-based fitness shaping, which replaces raw fitness by ordered utility weights, and specialized distribution parameterizations that avoid a literal general Fisher-matrix inversion.
⚠️ Curated Limitations & Negative Results
- The score identity and Fisher construction require suitable regularity, integrability and a usable parameterization. The empirical Fisher matrix can be singular or numerically unstable even when the population Fisher is invertible; a small batch does not automatically support the displayed inverse.
- Natural-gradient invariance describes the local direction under the appropriate exact metric. It is not an assertion that every finite, sampled parameter update is invariant under arbitrary reparameterization, stays in a valid distribution domain or improves expected fitness. Variant-specific parameterizations and learning rates matter.
- Rank-based shaping changes the weighting criterion; its update should not be described as an unbiased estimate of the raw expected-fitness gradient in general. A Gaussian search distribution, exponential covariance update and adaptation sampling are not mandatory features of every NES family member.
- Only the declared method and discussion pages were inspected. The record does not certify every displayed variant formula, an executable xNES/SNES implementation, universal benchmark superiority or a finite-budget global optimum.
Artificial Intelligence Ontology (AIO) Grounding
Explicit AIO Snapshot Assessment: No Match
Rationale: No inspected AIO target captures this general optimization procedure or a defining component. The assessment distinguishes evolutionary language models, learning/training tasks and predictive models from the native algorithm; it is bounded to this pinned vocabulary.
Search Scope: All 442 pinned AIO CLASS identities and scoped synonyms, all 270 preceding maintained identities, ignored/hidden checkout text, and semantic inspection of eight exact AIO targets plus seven neighboring native mechanisms.
| Examined AIO Candidate | Rejection Reason |
|---|---|
AIO:EvolutionaryLLM
|
The CLASS definition is a large language model with evolutionary optimization. Its broad related synonym evolutionary algorithms does not make a general continuous optimizer a language model or an exact vocabulary match. |
AIO:TrainingStrategy
|
The pinned definition concerns preprocessing/training of machine-learning models. General black-box continuous optimization is not inherently model training; an optional application does not establish a necessary mapping. |
AIO:MachineLearningTask
|
The target denotes building methods that learn from data. These algorithms can optimize arbitrary evaluated functions without a learning task, so shared optimization vocabulary alone is insufficient. |
AIO:MetaLearning
|
Adapting a search distribution or optimizer state does not require learning from metadata about machine-learning experiments as in the pinned definition. |
AIO:ActiveLearning
|
Candidate evaluation is not the defined task of querying labels for data points. |
AIO:ReinforcementLearning
|
A possible policy-search application does not turn the general optimizer into the pinned learning task; no environment/reward-policy interface is inherent to the full identity. |
AIO:RandomForest
|
Population members and search distributions are not a predictive decision-tree ensemble; no random-forest surrogate is required. |
AIO:EvaluationBias
|
This is a bias concept, not the optimizer. Possible evaluation or sampling problems are limitations rather than identity mappings. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Daan Wierstra, Tom Schaul, Tobias Glasmachers, Yi Sun, Jan Peters, Jürgen Schmidhuber
|
primary paper
full_text
|
sha256:80ec84951836…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T12:18:29Z
Updated: 2026-10-11T12:18:29Z
Source File:
data/curated/NaturalEvolutionStrategies.yaml