Evolutionary Algorithm¶

A population-based stochastic search family that repeatedly evaluates encoded candidates, selects parents or survivors, creates heritable variants, and replaces population members.

Version: v2 · 2026-09-06 · History
Domain-specific #: 1801
Origin domain: evolutionary computation
Subdomain: population based metaheuristics
Aliases: EA

Core Idea¶

An evolutionary algorithm is a family of population-based stochastic search and optimization procedures inspired by variation, differential selection, and inheritance. It maintains computationally represented candidate solutions, evaluates their quality or behavior, uses selection to bias which candidates reproduce or survive, creates offspring with representation-compatible variation operators, and updates the population. The cycle continues until a budget, target, convergence test, or other termination condition is met.

Eiben and Smith’s standard decomposition names representation, evaluation, population, parent selection, variation, survivor selection or replacement, initialization, and termination as the core components of an evolutionary algorithm.^[1] Their treatment also separates major variants—genetic algorithms, evolution strategies, evolutionary programming, and genetic programming—by representation and operator choices rather than treating one subtype as the entire family.^[2]

The biological language is operational, not a claim that a computer run reproduces natural evolution faithfully. “Individuals” are encoded candidate solutions; “fitness” is an evaluation used by the search; “reproduction” applies algorithmic operators; and “generations” or steady-state updates are computation schedules. The essential recurrence is evaluate → select → vary → replace, with information transmitted through offspring representations.

Recombination is common but not mandatory in every evolutionary family. Mutation or another heritable variation mechanism must introduce or modify candidates, and selection must make evaluation consequential. A population subjected only to independent random resampling is random search, not an evolutionary algorithm.

Structural Signature¶

Problem and search space: a declared task with candidate solutions that can be represented computationally.
Representation: a genotype, parameter vector, tree, permutation, rule set, graph, or other heritable encoding.
Population: multiple candidate instances maintained simultaneously or through an explicit parent–offspring pool.
Evaluation: a fitness, objective, ranking, behavior descriptor, constraint-handling rule, or interaction-based assessment.
Parent selection: a rule that chooses candidates to generate offspring, ordinarily biased by evaluation or diversity policy.
Variation: mutation and optionally recombination produce new encoded candidates from selected material.
Survivor selection or replacement: a rule determines which parents and offspring constitute the next population.
Termination: an evaluation budget, generation limit, target value, stagnation test, or other explicit stopping condition.

Recognition test. Trace one iteration and identify candidates, evaluation, selection, heritable variation, and population update. Evaluation must influence differential reproduction or retention, and variants must be generated from existing candidate material rather than drawn independently forever. If either link is absent, the evolutionary identity collapses.

What It Is Not¶

It is not biological evolution itself. Natural populations do not receive a designer-specified objective function or stop after a computational budget, and biological fitness is context-dependent reproductive success rather than simply a scalar engineering score. Evolutionary algorithms borrow a mechanism family and deliberately engineer it.

It is not synonymous with genetic algorithm. Genetic algorithms are a prominent subtype historically associated with chromosome-like encodings, crossover, and mutation. Evolution strategies emphasize real-valued parameters and mutation strategy; genetic programming evolves executable trees or programs; evolutionary programming developed through yet another operator tradition.

It is not every population-based metaheuristic. Particle swarm optimization updates particles through velocity-like attraction rules rather than parent selection and heritable offspring variation. Ant-colony optimization updates a shared pheromone model. Both may be bio-inspired and population-based without satisfying the EA reproduction cycle.

It is not random search, hill climbing, simulated annealing, or exhaustive enumeration. Those can share stochastic proposals, objective evaluations, or retention, but they do not necessarily maintain and reproduce an evaluated population.

Scope of Application¶

Evolutionary algorithms are used for black-box, discontinuous, noisy, mixed-variable, combinatorial, multimodal, and multiobjective problems where derivatives or exact solvers are unavailable or inconvenient. They also support design exploration, automated program construction, controller search, scheduling, feature selection, and model calibration.

The family includes generational and steady-state population models, single- and multiobjective evaluation, fixed and self-adaptive operators, constrained optimization, coevolution, interactive evaluation, neuroevolution, and quality-diversity methods. Each extension must still bind representation, evaluation, selection, variation, and population update.

An EA does not guarantee a global optimum merely because it explores stochastically. Performance depends on representation, operators, parameter settings, evaluation budget, and problem structure. A no-free-lunch boundary remains: success on a problem class reflects alignment between search bias and that class, not universal superiority.

Clarity¶

Representation and solution meaning must be separated. A genotype is the stored object variation operators manipulate; a phenotype is the decoded solution evaluated in the problem domain. They can coincide for a real vector, but tree grammars, indirect encodings, and repair functions make the mapping consequential.

Fitness is also not always the raw objective. Minimization may be converted to ranking; penalties or feasibility rules may handle constraints; multiobjective methods can use Pareto dominance and diversity; coevolution can score interaction outcomes; quality-diversity search can preserve behavioral niches. The stable role is an evaluation that affects selection, not one universal scalar formula.

Parent selection and survivor selection are distinct. The first chooses reproductive inputs. The second chooses the next population after offspring exist. Tournament selection, truncation, elitism, age-based replacement, and \((\mu,\lambda)\) or \((\mu+\lambda)\) strategies allocate these roles differently.^[1]

Manages Complexity¶

Large search spaces make enumeration impossible. An EA samples a bounded population, uses evaluations to concentrate future sampling, and preserves variation so search does not reduce immediately to one local trajectory. Recombination can combine material from several candidates; mutation explores nearby or novel regions.

The population carries parallel hypotheses. It can maintain alternative basins, trade-offs, niches, or building blocks and can return a set of solutions rather than one point. This distributed state is the family’s distinctive complexity-management resource.

The benefit has a cost: evaluation often dominates runtime, and stochastic outcomes require repeated runs and statistical reporting. Population size, variation strength, and selection pressure introduce a design space of their own. The abstraction replaces exhaustive search with biased sampling; it does not make complexity disappear.

Abstract Reasoning¶

Let \(P_t=\{x_1,\ldots,x_\mu\}\) be the population at iteration \(t\), and let \(E(x)\) be an evaluation. A generic cycle samples parents according to \(S_p(P_t,E)\), applies a variation kernel \(V(\cdot\mid\text{parents})\) to produce offspring \(O_t\), evaluates them, and uses survivor rule \(S_s(P_t,O_t,E)\) to form \(P_{t+1}\).

This reveals two coupled distributions. Selection reweights the current population toward favored candidates; variation spreads mass into new candidates. Excessive selection pressure with weak variation can collapse diversity prematurely. Excessive variation or weak selection approaches random sampling. Search depends on balancing exploitation and exploration.

Elitism—guaranteeing that selected high-quality candidates survive—can make the best-so-far objective nonworsening, but it does not imply the whole population improves or that the global optimum will be reached within a practical budget. Ergodic mutation plus infinite time can support asymptotic claims for some schemes, yet finite-run performance remains empirical and problem-dependent.

Knowledge Transfer¶

The EA cycle transfers literally across bit strings, real vectors, permutations, syntax trees, neural-network encodings, and rule sets. The representation-specific operators change, but evaluation-guided selection, inherited variation, and population replacement remain.

Transfer requires operator repair. One-point crossover that is valid for bit strings can destroy permutation validity; Gaussian mutation natural for real vectors is meaningless for syntax trees. A successful transfer preserves feasible representation semantics or adds repair and constraint handling.

The biological metaphor also transfers to nonoptimization goals such as novelty search and quality-diversity. Here evaluation may reward behavioral novelty or maintain an archive across niches rather than maximize a single objective. The algorithm remains evolutionary because evaluation still differentially shapes reproduction or retention.

Examples¶

Bit-string genetic algorithm. A population of length-\(d\) bit strings is scored by the number of one bits. Tournament selection chooses parents, crossover recombines segments, bit-flip mutation introduces variants, and elitist replacement forms the next generation. The complete EA roles are visible.

Real-valued evolution strategy. Candidate vectors and mutation step sizes are evolved together. Parents generate Gaussian-mutated offspring, the offspring are evaluated on a continuous objective, and the best \(\mu\) of \(\lambda\) offspring survive. Recombination may be optional while selection, mutation, and population replacement remain.

Genetic programming. Individuals are syntax trees representing programs. Fitness comes from task performance; subtree crossover and mutation create offspring; type or grammar constraints preserve valid programs. The candidate substrate changes without changing the cycle.

Nonexample. Particle swarm optimization maintains many candidates and evaluates them, but particles move by personal- and neighborhood-best velocity updates. Without parent selection and heritable offspring variation, it belongs to a neighboring metaheuristic family rather than EA proper.^[3]

Structural Tensions¶

Selection pressure versus diversity: strong preference accelerates exploitation but can erase alternatives. Diagnostic: track genotype or phenotype diversity alongside best fitness.
Variation reach versus heritability: disruptive operators explore broadly but may destroy useful structure. Diagnostic: measure offspring feasibility and parent–offspring similarity under each operator.
Fitness fidelity versus evaluation cost: accurate scoring can dominate the budget. Diagnostic: report the fraction of runtime and uncertainty attributable to evaluations and validate any surrogate.
Representation generality versus operator validity: a portable algorithm template can produce invalid candidates in a new encoding. Diagnostic: prove closure or quantify repair and rejection rates.
Elitism versus adaptability: preserving incumbents prevents regression but can entrench solutions in changing objectives. Diagnostic: compare best-so-far retention with response after an environmental shift.
Stochastic exploration versus reproducibility: randomness diversifies search but makes one run unreliable evidence. Diagnostic: record seeds, budgets, distributions across independent runs, and baseline comparisons.
Biological analogy versus computational purpose: metaphor can import false claims about natural selection. Diagnostic: restate every role as an executable data structure or operation.

Structural–Framed Character¶

Evolutionary Algorithm is structurally defined by a recurring computational loop, yet its framing comes from evolutionary computation. Population, selection, variation, and inheritance have exact algorithmic roles. The biological terms guide design but do not license a claim of biological realism.

Different subfamilies can omit a familiar surface feature such as crossover while preserving the structural identity. What cannot disappear is evaluation-guided differential retention or reproduction coupled to heritable candidate variation.

Structural Core vs. Domain Accent¶

The portable core is variation, selection, and retention. The indispensable domain accent adds digital candidate representations, an executable evaluation, explicit parent and survivor mechanisms, representation-aware variation, population state, resource budgets, and termination.

The existing Natural Selection and Variation Strategies primes capture broader recurrence across substrates. Evolutionary Algorithm is not another prime because its literal examples remain computational search procedures; it is a stable domain-specific implementation family.

prime:algorithm is the proposed minimal parent by strict specialization. An EA is an executable problem-solving procedure with defined state and repeated operations. The child adds a population, evaluation-guided selection, heritable variation, and replacement.

prime:natural_selection supplies the variation–selection–retention engine and is a strong related abstraction, but it is substrate-neutral and does not entail engineered representation or termination. prime:optimization names a common purpose, yet some EAs pursue novelty or diverse repertoires. prime:variation_strategies supplies deliberate generation of alternatives but not differential population replacement.

Relationships to Other Abstractions¶

Current abstraction Evolutionary Algorithm Domain-specific

Parents (1) — more general patterns this builds on

Evolutionary Algorithm is a kind of Algorithm Prime

prime:algorithm is the proposed minimal parent by strict specialization.

Hierarchy paths (2) — routes to 2 parentless roots

Evolutionary Algorithm → Algorithm → Function (Mapping)

Show alternative path (1)

Neighborhood in Abstraction Space¶

Evolutionary Algorithm sits in a sparse region of the domain-specific corpus (84^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Population Genetics & Selection (15 abstractions)

Nearest neighbors

Premature convergence — 0.82
Evolutionary acquisition of neural topologies — 0.81
Evolutionary Attractor — 0.81
Population-based incremental learning — 0.80
Evolutionary data mining — 0.80

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Genetic algorithm: one major EA subtype, not the entire family.
Evolution strategy: a real-valued, mutation-centered EA tradition.
Genetic programming: EAs whose candidates are programs or expression structures.
Natural selection: biological or substrate-neutral differential reproduction, not necessarily an algorithm.
Particle swarm optimization: population-based motion through social attraction, not offspring variation.
Ant-colony optimization: search through a shared pheromone model.
Random search: independent sampling without evaluation-guided heredity.
Hill climbing: typically one-current-state local improvement without a reproducing population.
Evolutionary computation: the broader research field containing EAs and related methods.

References¶

[1] A. E. Eiben and J. E. Smith, “What Is an Evolutionary Algorithm?”, in Introduction to Evolutionary Computing, Springer, first-edition chapter PDF, https://www.cs.vu.nl/~gusz/ecbook/Eiben-Smith-Intro2EC-Ch2.pdf. registry ↩a ↩b

[2] A. E. Eiben and J. E. Smith, Introduction to Evolutionary Computing, 2^nd ed., Springer, 2015, https://doi.org/10.1007/978-3-662-44874-8. registry ↩

[3] Sean Luke, Essentials of Metaheuristics, 2^nd ed., George Mason University, 2013, https://people.cs.gmu.edu/~sean/book/metaheuristics/Essentials.pdf. registry ↩