Overview

Evolution strategies (ES) are a family of stochastic optimization techniques that mimic natural evolution to search for high-quality solutions in continuous or mixed spaces. They belong to the broader field of evolutionary algorithms and emphasize mutation-driven variation, selection of better individuals, and adaptation of search parameters over time.

Key components and variants

An ES typically maintains a population of candidate solutions represented as real-valued vectors. New candidates are produced by applying variation operators—primarily Gaussian mutation—and sometimes recombination. Selection chooses which individuals survive to the next generation. Common notations describe population structure, for example (1+1)-ES, (μ,λ)-ES and (μ+λ)-ES; these indicate parent and offspring counts and whether parents compete with offspring.

  • Mutation: normally additive, drawn from a distribution with adjustable scale (step size).
  • Recombination: averaging or discrete mixing of parents to form offspring.
  • Selection: deterministic or stochastic ranking by objective value.
  • Self-adaptation: strategy parameters such as step sizes are encoded and evolved along with solutions.

History and development

The term "evolution strategy" was introduced in Germany by researchers including Ingo Rechenberg and later Hans-Paul Schwefel as part of early work on using randomized search to solve engineering problems. Over subsequent decades the approach evolved; significant later advances include covariance matrix adaptation (CMA-ES), which adapts the full mutation covariance to better follow the landscape.

Applications and strengths

ES methods are widely used for continuous black-box optimization where derivatives are unavailable or unreliable. Typical applications include parameter tuning in machine learning, control and robotics, engineering design, and simulation-based optimization. They are robust to noise and multimodality and scale well with moderate dimensionality when properly configured.

Distinctions and notable facts

Compared with genetic algorithms, evolution strategies usually operate on real-valued vectors and prioritize mutation and parameter self-adaptation over crossover. Practical rules and mechanisms—such as the one-fifth success rule for step-size control and more modern cumulative step-size or covariance adaptation methods—help balance exploration and exploitation. Variants range from simple single-individual strategies to sophisticated population-based algorithms used in research and industry.