Must the objective be differentiable?

No; its evaluations must be available, while the chosen distribution supports the needed calculations.

Is NES identical to a genetic algorithm?

No; its defining update uses a search distribution and natural-gradient principles.

Does it guarantee the global optimum?

No general guarantee follows from the method's name or construction.

AI Glossary / Probabilistic machine learning / Advanced

Natural evolution strategy

Definition

A natural evolution strategy is a member of a family of optimization methods that improve a parameterized search distribution using a natural gradient. Candidates are sampled, evaluated, and used to adjust that distribution toward higher expected fitness. The method can optimize a black-box objective without requiring derivatives of the objective with respect to the candidates themselves.

What it means

Imagine tuning several controls in a simulator that returns a performance score but provides no useful gradient. Instead of maintaining only one proposed setting, a natural evolution strategy maintains a distribution of possible settings. Its location determines where to search, while its spread determines how broadly to explore.
Natural refers to the geometry used to adjust the distribution. Equal numerical changes in two distribution parameters may have very different effects on the probabilities of sampled candidates. The natural gradient accounts for this sensitivity through Fisher information.

How it works

Choose a search distribution, such as a multivariate normal, and initialize its parameters. Draw a population of candidates and evaluate their fitness. The basic search-gradient identity uses the average of fitness multiplied by the derivative of the search distribution's log density. This differentiates the distribution rather than the black-box evaluation procedure.
Precondition that estimated gradient using the inverse Fisher information matrix, or an equivalent efficient calculation, and update the distribution parameters with a step. Repeat sampling and evaluation until a stopping criterion is reached. In Gaussian versions, adapting both the mean and covariance allows the search to change its preferred region and directions.
Practical variants differ in distribution structure and update implementation. A full covariance can represent relationships between search directions but costs more than a diagonal representation. Rank-based fitness shaping and other refinements may replace the raw-fitness weighting used in the basic derivation. Finite populations produce noisy estimates, and a large update can still reduce performance.

Three practical examples

A robotics team tunes controller parameters in a simulator whose score includes discontinuous success and failure events. Candidate simulations can be evaluated separately, while final controller performance requires testing beyond the scenarios used during search.
A design team optimizes continuous settings of a component using an external engineering evaluator. A natural evolution strategy can use returned scores without accessing the evaluator's derivatives. Constraints and invalid configurations still need an explicit treatment.
A software team experiments with continuous parameters of a decision policy inside a replay environment. Parallel evaluations may shorten elapsed search time, but the number and cost of evaluations still determine whether this approach is economical.

Think of it this way

Picture moving and reshaping a cloud of trial solutions. The evaluation tells you which parts of the cloud look promising. The natural-gradient adjustment accounts for how sensitive that cloud is to its own controls, rather than assuming that one numerical unit of change means the same thing for every control.

Worked example

Consider the teaching objective f(x) = −(x − 3)², which we want to maximize. Use a normal search distribution with mean mu and fixed variance 4. We adapt only the mean here to isolate the calculation; this is not a complete covariance-adapting NES implementation.
Start at mu = 0 and suppose two illustrative candidate values are −2 and 2. Their fitnesses are −25 and −1. The derivative of the normal log density with respect to mu is (x − mu)/4, giving scores −0.5 and 0.5. The raw-fitness gradient estimate is [−25 × (−0.5) + (−1) × 0.5]/2 = 6.
For a normal mean with fixed variance 4, Fisher information is 1/4. The natural-gradient estimate is therefore 6/(1/4) = 24. With a learning rate of 0.05, the new mean is 0 + 0.05 × 24 = 1.2. The variance stays at 4 by the example's explicit restriction.
Because this toy objective is known analytically, we can check expected fitness: J(mu) = −[(mu − 3)² + 4]. It changes from −13 at mu = 0 to −7.24 at mu = 1.2, an improvement for this step. The update itself used candidate evaluations and distribution derivatives, not the objective's derivative.
Two candidates do not establish a reliable general optimization procedure. This particular estimate happens to equal the exact mean gradient at the starting point; that agreement is not guaranteed for other samples. The best possible expected fitness with variance held at 4 is −4, showing why the search distribution's spread matters even when its mean reaches the optimum x = 3.

Illustrative case study

Fictional company Juniper Motion proposes tuning a simulated conveyor controller using a natural evolution strategy. Its engineers define a score that combines delivery completion with penalties for unacceptable behavior and compare search methods under the same evaluation budget. They repeat experiments across seeds and reserve different load patterns for subsequent testing.
Before any operational decision, they examine individual candidate behavior rather than relying only on the aggregate score. A controller can exploit weaknesses in an evaluator without solving the intended problem. This invented scenario describes an evaluation plan; it does not claim better control, faster optimization, or successful deployment.

Common misconceptions and mistakes

Derivative-free evaluation does not mean that no gradients appear anywhere: the method uses derivatives of its search distribution. Nor is every evolutionary or random-search algorithm a natural evolution strategy.
Parallel candidate evaluation reduces elapsed time only when resources and evaluator behavior permit it; it does not make evaluations free. Prematurely shrinking exploration can trap the search, while excessively broad sampling can waste the budget. The natural gradient addresses parameter sensitivity but does not remove sampling noise, local optima, misleading fitness definitions, or finite-step tuning decisions.

Business use

Consider this family when a valuable continuous optimization problem offers evaluations but lacks practical objective gradients. Compare evaluation cost, repeatability, and final performance against appropriate alternatives. Budget for validation outside the search environment, and record the chosen search distribution, constraints, stopping rules, and performance variation across runs.

Check your understanding

Why divide the estimated gradient by 1/4?

Answer: That is the Fisher information for this mean parameter.

What becomes 1.2?

Answer: The search mean.

Why is expected fitness still below zero at mean 3?

Answer: The fixed search variance continues producing non-optimal candidates.

Related terms

Fitness function
Evolution strategy
natural gradient
Fisher information
black-box optimization
search distribution
covariance matrix adaptation

Learn next

Study score-function gradient estimates, then compare full-covariance and separable search distributions under a fixed evaluation budget.

Sources

Wierstra et al., Natural Evolution Strategies (2014), Sections 2–3 independently read: https://www.jmlr.org/papers/volume15/wierstra14a/wierstra14a.pdf . The fixed-variance calculation and Juniper Motion scenario are original illustrations.

Previous: Nats Next: Natural image statistics