Does an evolution strategy require a population of many candidates?

Not always. Simple variants use one parent and one offspring, while others use larger populations.

Is CMA-ES the whole field?

No. It is one important strategy with a particular distribution-adaptation mechanism.

Does avoiding derivatives make these methods universally better?

No. Derivative-based methods can be more efficient when the objective and gradients suit them.

AI Glossary / Graphs and symbolic AI / Intermediate

Evolutionary strategies

Definition

Evolutionary strategies, commonly called evolution strategies or ES, are optimization methods that improve candidate solutions through variation, evaluation, and selection. They are especially associated with real-valued parameters and with adapting how new candidates are generated. The aim is to find useful solutions by evaluating an objective, often without needing its derivatives.

What it means

A candidate might be a vector of settings for a simulation. The objective assigns a score to that vector, such as an error to minimize. The strategy generates modified candidates, evaluates them, and uses the better results to influence the next search step.

The word evolutionary describes the algorithmic use of variation and selection. It does not mean that the system is alive, understands its purpose, or necessarily changes its software structure.

Different evolution strategies use different population sizes, selection rules, and adaptation mechanisms. CMA-ES, or covariance matrix adaptation evolution strategy, is a prominent member of the family. It adapts the shape and scale of a sampling distribution so that future trials can reflect promising directions discovered during search.

How it works

Define a representation, an objective, valid parameter ranges, and an evaluation budget. Initialize one or more candidate solutions or a distribution from which candidates will be sampled.

Generate offspring by random variation, often using Gaussian perturbations in continuous spaces. Evaluate the candidates and apply the chosen selection rule. Some schemes allow existing parents to survive; others select the next generation from offspring only.

The strategy can adjust mutation step sizes or other distribution parameters based on successful search behavior. In population-based variants, information from several selected candidates may be recombined. Adaptation aims to make subsequent trials more effective, but its details are algorithm-specific.

Continue until a budget or stopping condition is reached. Retain the best evaluated feasible result and validate it independently where appropriate. A good score on a simulation or training objective is evidence about that objective, not an automatic guarantee of real-world performance.

Three practical examples

1. An engineer searches for simulation parameters that reduce the discrepancy between simulated and observed behavior when a usable gradient is unavailable.

2. A design team varies continuous shape parameters and evaluates each candidate through a computational model.

3. A researcher tunes a small vector of controller settings in a simulator, comparing repeated runs and checking the selected settings under separate test conditions.

Think of it this way

Think of making several slightly different prototypes, testing them, and using the better versions to guide the next round. The process learns which changes are worth exploring.

The analogy has limits. An evolution strategy follows mathematical sampling and selection rules, and its objective may omit qualities a human designer notices. Repeatedly improving the score does not establish that the score captures everything that matters.

Worked example

Consider minimizing f(x) = (x − 3)² with a simple illustrative one-parent, one-offspring strategy. Start at x = 0, where the score is 9. Keep the parent unless the proposed offspring has a lower score.

Suppose the first sampled perturbation is +1.5. The offspring is x = 1.5 with score 2.25, so it replaces the parent. A subsequent perturbation of −1 produces x = 0.5 with score 6.25, which is rejected. The retained parent remains 1.5.

Next, a perturbation of +1 produces x = 2.5 with score 0.25, so it is accepted. A following perturbation of +1 produces x = 3.5, also scoring 0.25. Under the stated strict-improvement rule, this tied proposal is rejected.

The example fixes a sequence of proposed perturbations to make evaluation and selection transparent. It does not implement a full step-size adaptation policy or CMA-ES. The known optimum x = 3 helps us inspect the calculation, but the search decisions use only the evaluated scores.

Illustrative case study

Fictional case: Amber Acoustics calibrates a simplified room simulator using several continuous parameters. Its team considers an evolution strategy because the simulation is available as an evaluator but does not provide useful derivatives.

The team caps the number of evaluations, repeats the search with different random seeds, and tests the resulting settings against recordings excluded from calibration. It discovers that a lower calibration error can coexist with poor results on another room layout. The lesson is to validate the objective and test coverage alongside the optimization method.

Common misconceptions and mistakes

Do not equate evolution strategies with every evolutionary algorithm. Genetic algorithms and other approaches can share variation and selection while using different representations and mechanisms.

Do not assume derivative-free means evaluation-free. Many objective evaluations may be needed, which matters when simulations are expensive.

Avoid treating the best observed candidate as a proven global optimum. Search budgets, local structure, and randomness limit that conclusion.

Do not silently ignore bounds or feasibility restrictions. Constraint handling is part of the optimization design, and a numerically attractive but invalid candidate may be unusable.

Business use

Evolution strategies can be useful when a business has a reliable evaluator for continuous settings but lacks convenient derivatives. Suitability depends on evaluation cost, dimension, noise, constraints, and available alternatives.

Compare methods using the same evaluation budget and representative cases. Report distributions across runs instead of one fortunate result. Keep a record of parameter scaling, initialization, random seeds, and objective definitions so that promising configurations can be reproduced and assessed.

Check your understanding

In the worked example, should the proposal x = 0.5 replace the retained x = 1.5?

Answer: No. Its score is 6.25 rather than 2.25, so it is worse for the stated minimization objective and selection rule.

Related terms

Learn next

Study the difference between parent-plus-offspring and offspring-only selection, then examine how mutation scale affects exploration. Follow with a basic CMA-ES explanation of mean, covariance, and step size. Keep performance claims tied to evaluated problem classes.

Sources

Primary sources: Nikolaus Hansen, The CMA Evolution Strategy: A Tutorial: https://arxiv.org/pdf/1604.00772

Author's overview and bibliography: https://www.cmap.polytechnique.fr/~nikolaus.hansen/cmaesintro.html

Scope: the supplied term is interpreted as the evolution-strategy optimization family. The scalar selection example and fictional case are original; the example is not a complete CMA-ES implementation.

Previous: Evolutionary robotics Next: Evolvability (computer science)