{
  "id": "genetic-algorithm",
  "title": "Genetic Algorithm",
  "lang": "en",
  "source_lang": "en",
  "summary": "A genetic algorithm is a population-based optimization method that evolves candidate solutions through selection, recombination, and mutation.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Genetic search",
    "GA"
  ],
  "generated_at": "2026-11-02",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/genetic-algorithm/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/genetic-algorithm/",
      "json": "https://aiwiki.page/zh/genetic-algorithm.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A genetic algorithm (GA) is a population-based randomized algorithm for searching a space of candidate solutions. It repeatedly evaluates candidates, preferentially selects some of them as parents, and produces new candidates through recombination and mutation. The method borrows terminology from biological evolution, but its operations are mathematical and computational constructions rather than detailed simulations of genetics. Genetic algorithms belong to the broader family of evolutionary computation and are commonly used for mathematical optimization, design exploration, and automated search.\n\nTheir distinctive feature is the interaction between a population, selection based on performance, and variation that generates new combinations of solution components. This makes them applicable to discrete, continuous, and mixed search spaces, including problems whose objectives are discontinuous, noisy, expensive to evaluate, or unavailable in differentiable form. They do not generally guarantee an optimal answer within a practical computational budget. Their effectiveness depends on how solutions are represented, how variation respects the problem’s structure, and how evaluation and selection balance progress with the preservation of useful diversity."
    },
    {
      "heading": "Historical development and intellectual foundations",
      "text": "Attempts to use evolutionary ideas in computation preceded the modern genetic algorithm. In 1954, Nils Aall Barricelli conducted numerical experiments on self-reproducing structures at the Institute for Advanced Study in Princeton. During the late 1950s and early 1960s, researchers explored artificial evolution as a mechanism for adaptation and automatic problem solving. These projects differed substantially in representation and purpose, but they established the possibility of studying selection and reproduction as computational processes rather than merely as biological analogies.\n\nThe research tradition most directly associated with genetic algorithms developed around John H. Holland at the University of Michigan. Holland investigated how adaptive systems could discover useful combinations of interacting components under uncertain conditions. His 1975 book, Adaptation in Natural and Artificial Systems, provided a foundational account of genetic search, including binary-string representations, crossover, and the analysis of patterns called schemata. The emphasis was broader than numerical optimization: genetic algorithms were intended as models and mechanisms of adaptation in complex environments.\n\nSeveral neighboring traditions developed independently. In Germany, Ingo Rechenberg and Hans-Paul Schwefel developed evolution strategies during the 1960s, initially in connection with experimental engineering optimization. Lawrence J. Fogel, Alvin J. Owens, and Michael J. Walsh described evolutionary programming in their 1966 book Artificial Intelligence through Simulated Evolution, emphasizing the evolution of finite-state predictive systems. These approaches later became grouped under evolutionary computation. Their historical distinctions involved representation, variation operators, and selection, although contemporary implementations often combine techniques across these boundaries.\n\nGenetic algorithms gained a wider engineering and computing audience during the 1980s and 1990s. David E. Goldberg’s 1989 book, Genetic Algorithms in Search, Optimization, and Machine Learning, presented their mechanisms and applications in an accessible technical form. Research expanded into scheduling, design optimization, classifier systems, and multiobjective search. John R. Koza’s 1992 book on genetic programming helped establish the evolution of executable programs as a prominent related field.\n\nThe field’s development also changed its explanatory emphasis. Early discussions often stressed the propagation and recombination of short, favorable patterns in chromosomes. Later research gave greater attention to problem structure, interactions among variables, population dynamics, runtime analysis, and controlled experimentation. Modern genetic algorithms are consequently better understood as a flexible family of search methods than as one fixed procedure with a universal encoding or parameter set."
    },
    {
      "heading": "Optimization model and evolutionary cycle",
      "text": "A genetic algorithm operates on a search space \\(\\mathcal{X}\\) and an objective that assigns quality to candidate solutions. For a maximization problem, the formal goal is\n\n\\[x^\\star \\in \\operatorname*{arg\\,max}_{x\\in\\mathcal{X}} f(x).\\]\n\nFor minimization, the direction of comparison is reversed. The objective might measure manufacturing cost, route length, predictive error, energy consumption, or agreement with experimental observations. Constraints restrict which candidates are permissible. Although the biological term fitness is often used for \\(f(x)\\), the quantity used by the selection mechanism need not equal the original objective: it may incorporate ranking, penalties, normalization, or several criteria.\n\nThe population at generation \\(t\\) can be written as \\(P_t=(x_1,\\ldots,x_N)\\), where \\(N\\) is the population size. Initialization creates the first candidates, often randomly but sometimes using feasible constructions, existing designs, or solutions from another heuristic. Each candidate is evaluated. Parent selection then samples candidates according to a rule that favors particular qualities. Recombination creates offspring by combining parental information, while mutation changes components of an offspring or creates further variation independently.\n\nReplacement determines the next population. In a generational algorithm, a batch of offspring replaces most or all of the current population. In a steady-state algorithm, only a few candidates are inserted at a time. Elitism explicitly preserves one or more of the best candidates found so far. These design choices are independent of the particular crossover operator: two algorithms can use identical variation but behave differently because their survival and replacement rules differ.\n\nA simplified generational procedure can be expressed as pseudocode:\n\nP ← initialize N candidates\nevaluate every candidate in P\nrecord the best candidate found\n\nwhile the stopping condition is not met:\n    Q ← empty offspring population\n    while Q is not full:\n        parents ← select from P\n        offspring ← recombine parents, or copy selected parents\n        offspring ← mutate offspring\n        evaluate offspring\n        add offspring to Q\n    P ← select survivors from P and Q\n    update the best candidate found\n\nreturn the recorded result\n\nA run typically stops after a fixed number of evaluations, a time limit, attainment of a target value, or a prolonged lack of improvement. None of these conditions alone certifies global optimality. The best-so-far result is therefore usually an approximation accompanied by its objective value, feasibility status, and computational cost. For multiobjective problems, the output may instead be a collection of trade-off solutions. The evolutionary cycle is simple, but its behavior depends on the detailed interaction of evaluation, representation, selection, variation, and replacement."
    },
    {
      "heading": "Representation, decoding, and constraints",
      "text": "The representation defines what an individual is and what changes the algorithm can make to it. In a classical binary genetic algorithm, an individual is a string of bits. A string can directly encode a subset, with each position indicating whether an item is included, or indirectly encode numerical variables. A \\(k\\)-bit unsigned integer \\(b\\), for example, can represent a value in \\([a,c]\\) through\n\n\\[x=a+\\frac{b}{2^k-1}(c-a).\\]\n\nThis introduces a finite resolution determined by \\(k\\). Binary encoding is not obligatory: real-valued vectors, permutations, trees, graphs, and mixed structures are all used when they better match the problem.\n\nThe distinction between genotype and phenotype separates the stored representation from the solution that is evaluated. A genotype may specify priorities that a decoder converts into a feasible schedule rather than directly listing job start times. Several genotypes may decode to the same phenotype, creating redundancy. Conversely, a small genotypic change can cause a large phenotypic change. These relationships matter because the algorithm explores the neighborhoods induced by its operators, not merely the abstract set of feasible solutions.\n\nA useful representation must work with its variation operators. For the traveling salesperson problem, a tour can be represented as a permutation of cities. Ordinary substring exchange can create repeated cities and omit others, so specialized crossover is needed. An alternative encoding uses random numerical keys whose sorted order determines a permutation. This permits conventional real-valued operations, although the geometry of the key space differs from the geometry of route changes. There is no representation that is uniformly best independent of the operators and objective.\n\nConstraints can be enforced in several ways. A decoder can generate feasible solutions by construction; a repair operator can modify invalid offspring; or a penalty can reduce their selection quality. For minimization, a simple penalized objective is\n\n\\[F(x)=f(x)+\\lambda V(x),\\]\n\nwhere \\(V(x)\\geq 0\\) measures violation and \\(\\lambda\\) controls the penalty. Poor scaling can make infeasible candidates too attractive or suppress potentially useful exploration near constraint boundaries. Feasibility-based comparison offers another approach: feasible candidates outrank infeasible ones, feasible candidates are compared by objective, and infeasible candidates by violation.\n\nConstraint handling also changes the search distribution. Repair can repeatedly map many offspring to the same region, while rejecting infeasible individuals may become extremely inefficient when feasible solutions occupy a small fraction of the encoded space. Consequently, encoding and feasibility are not preliminary bookkeeping choices. They determine which structures can be inherited, how frequently useful candidates are generated, and whether the algorithm spends most of its budget evaluating meaningful solutions."
    },
    {
      "heading": "Selection, replacement, and population diversity",
      "text": "Selection controls how strongly observed performance influences reproduction. In fitness-proportionate selection, sometimes called roulette-wheel selection, candidate \\(i\\) receives probability\n\n\\[p_i=\\frac{w_i}{\\sum_{j=1}^{N}w_j},\\]\n\nwhere all weights \\(w_i\\) must be nonnegative and the denominator must be positive. These weights may be transformed objective values rather than raw measurements. Direct use of raw fitness can be problematic when values are negative, nearly equal, or dominated by one extreme individual. It also makes the algorithm sensitive to transformations that leave the underlying optimization problem unchanged.\n\nRank-based selection assigns probabilities according to ordering rather than numerical differences. Tournament selection samples a group of candidates and chooses its best member, possibly with an additional probabilistic rule. Increasing tournament size generally raises selection pressure: strong candidates reproduce more frequently, and weaker candidates disappear faster. Under deterministic tournaments with sampling with replacement and a unique best individual, the probability that a tournament of size \\(k\\) includes that individual is \\(1-(1-1/N)^k\\).\n\nReplacement creates a separate source of selection pressure. A candidate can be frequently chosen as a parent but still fail to survive if offspring compete strongly for population positions. Elitism protects high-quality discoveries from accidental loss, and with deterministic evaluations it can make the recorded best value monotonic. However, preserving too many elite copies can accelerate loss of diversity. A population containing many copies of one good solution may search less effectively than a population containing several distinct, moderately good solutions.\n\nPopulation diversity has several meanings. Genotypic diversity concerns differences among encoded candidates; phenotypic diversity concerns differences among decoded solutions; behavioral diversity concerns what candidates do when evaluated. A large spread of encoded strings need not imply meaningful differences in behavior. For binary populations, average pairwise Hamming distance or allele frequencies provide simple indicators. For structured designs, distances may need to reflect geometry, schedules, program behavior, or domain-specific characteristics.\n\nPremature convergence occurs when the population loses the variation needed for further progress before reaching a satisfactory region. Remedies include reducing selection pressure, introducing immigrants, limiting duplicates, or using niches in which similar candidates compete more directly. Fitness sharing reduces the effective reproductive advantage of candidates in crowded regions. Crowding methods often make offspring compete against similar individuals rather than against the entire population. These devices are useful only relative to an appropriate similarity measure.\n\nPopulation size is therefore not merely a measure of computational effort. It affects the number of alternatives maintained simultaneously and the reliability with which selection distinguishes useful structures. Larger populations can preserve diversity, but under a fixed evaluation budget they also permit fewer generations. The resulting trade-off depends on how much progress requires broad sampling versus repeated refinement."
    },
    {
      "heading": "Recombination, mutation, and interacting variables",
      "text": "Recombination, usually called crossover, creates offspring using information from two or more parents. In one-point crossover on equal-length strings, a cut position is selected and the suffixes are exchanged. Two-point crossover exchanges an interior segment. Uniform crossover decides independently at each position which parent supplies the value. These operators preserve different kinds of structure: one-point crossover tends to preserve contiguous blocks, whereas uniform crossover treats physical adjacency in the representation as less important.\n\nFor a binary string \\(x=(x_1,\\ldots,x_n)\\), standard bit mutation independently flips each position with probability \\(p_m\\). The expected number of flipped bits is \\(np_m\\). Setting \\(p_m=1/n\\) gives one expected flip per offspring, but it does not force exactly one flip: some offspring remain unchanged, while others change several positions. This rate is a common reference point, not a universal optimum. Mutation rates that are too small can impede exploration; rates that are too large can destroy inherited structure.\n\nReal-valued representations use different operators. Gaussian mutation adds a random perturbation, such as \\(x'_j=x_j+\\epsilon_j\\), with \\(\\epsilon_j\\) drawn from a zero-mean normal distribution. Its scale determines the characteristic step size. Arithmetic recombination forms weighted combinations of parental vectors. Simulated binary crossover, introduced by Kalyanmoy Deb and Ram Bhushan Agrawal in 1995, generates real-valued offspring using a distribution designed to reproduce certain properties of binary crossover. Boundary handling is necessary when such operations leave the permitted domain.\n\nFor permutations, mutation may swap two positions, insert an item elsewhere, or reverse a subsequence. Partially mapped crossover and order crossover preserve permutation validity through explicit rules for resolving inherited positions and order. Edge-oriented operators for routing problems emphasize parental adjacencies because an edge can be more important than a city’s absolute position. Thus, effective recombination depends on identifying what constitutes a reusable partial solution.\n\nThe difficulty of combining partial solutions is closely related to epistasis: the effect of one variable depends on the values of others. Two components that are individually favorable in their parental contexts may perform poorly when brought together. Linkage-aware methods attempt to detect or preserve interacting variable groups. Encoding strongly interacting variables in a way that operators frequently separate can make crossover destructive rather than productive.\n\nCrossover and mutation are complementary mechanisms, but their relative importance varies. Mutation can introduce values absent from the current population, while ordinary crossover can only rearrange existing parental information at the affected positions. Conversely, recombination can unite structures that mutation alone would assemble slowly. Some evolutionary algorithms operate successfully without crossover, so improvement cannot automatically be attributed to recombination merely because an implementation includes it. Establishing the contribution of an operator requires comparison against suitable variants under comparable evaluation budgets."
    },
    {
      "heading": "Theoretical analysis and its limits",
      "text": "The most familiar early theory of genetic algorithms is schema theory. In a binary representation, a schema is a pattern over \\(\\{0,1,*\\}\\), where an asterisk denotes an unspecified position. The schema \\(1*0*\\), for example, describes all four-bit strings whose first bit is 1 and third bit is 0. Its order \\(o(H)\\) is the number of specified positions, and its defining length \\(\\delta(H)\\) is the distance between the first and last specified positions.\n\nFor a simplified genetic algorithm using fitness-proportionate selection, one-point crossover, and independent bit mutation, a standard lower bound has the form\n\n\\[\\mathbb{E}[m(H,t+1)\\mid P_t] \\geq m(H,t)\\frac{\\bar f_H(t)}{\\bar f(t)} \\left(1-p_c\\frac{\\delta(H)}{n-1}\\right) (1-p_m)^{o(H)},\\]\n\nwhere \\(m(H,t)\\) is the number of instances of schema \\(H\\), \\(\\bar f_H(t)\\) is their mean fitness, \\(\\bar f(t)\\) is population mean fitness, \\(p_c\\) is crossover probability, and \\(n\\) is string length. The bound counts ways in which existing instances survive and ignores additional instances created by variation. It requires the selection assumptions and fitness conditions used in its derivation.\n\nThe theorem suggests that short, low-order patterns with above-average fitness can increase in expected frequency. It does not establish that such patterns combine into a global optimum, that their fitness advantage remains constant, or that an individual finite run follows the expectation closely. The associated building-block hypothesis is a proposed explanation of successful recombination, not a general proof of efficiency. Deceptive objective functions illustrate how locally favorable patterns can direct search away from the global optimum.\n\nA different approach models the population as a state in a Markov chain. In a finite search space, mutation that gives every candidate a positive probability of being generated can ensure eventual discovery of a global optimum with probability one under appropriate continuing-search assumptions. Elitism can then preserve that discovery. Such asymptotic statements say little about practical runtime: the probability of reaching the optimum may be so small that the expected waiting time is enormous.\n\nRuntime analysis instead studies quantities such as the expected number of evaluations needed to reach a target. Results are usually obtained for precisely specified algorithms and objective classes. They cannot automatically be transferred to all methods called genetic algorithms. Population size, crossover rules, selection intensity, and the objective’s structure can each alter the result substantially.\n\nThe no-free-lunch theorems for optimization further limit universal claims. Under specified assumptions, averaging equally over all functions on a finite domain removes the advantage of one non-revisiting black-box search method over another. Practical tasks are not sampled uniformly from all possible functions, so the result does not imply that algorithms are equally useful in applications. It instead emphasizes that successful search exploits structure in a relevant problem class."
    },
    {
      "heading": "Multiobjective, hybrid, and related methods",
      "text": "Many design problems involve objectives that conflict: lower cost may require greater energy use, while greater accuracy may require more computation. A multiobjective genetic algorithm evaluates each candidate using a vector \\(\\mathbf f(x)=(f_1(x),\\ldots,f_m(x))\\). For minimization, \\(x\\) dominates \\(y\\) if it is no worse in every objective and strictly better in at least one. The Pareto-optimal set contains feasible solutions not dominated by another feasible solution. Its image in objective space is the Pareto front.\n\nA population naturally accommodates multiple trade-offs, but ordinary scalar selection does not automatically preserve them. One approach combines objectives into a weighted sum. This is simple, yet requires weights and can miss portions of a nonconvex Pareto front. Pareto-based algorithms instead use dominance relations together with a diversity mechanism. Their purpose is generally to approximate a set of alternatives, not to declare one objectively best compromise without additional preferences.\n\nNSGA-II, introduced by Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan in 2002, combines elitist population merging, nondominated sorting, and crowding distance. Candidates are organized into dominance levels, and crowding distance favors less crowded portions of a front when a selection boundary is reached. Its widely used original sorting procedure has complexity \\(O(mN^2)\\) for \\(m\\) objectives and a population of size \\(N\\). With many objectives, dominance becomes less discriminating because numerous candidates are mutually nondominated; reference-direction and decomposition methods address aspects of this problem.\n\nHybrid algorithms combine evolutionary exploration with other search methods. A memetic algorithm commonly applies local search to some or all individuals, using the population to explore different basins while local improvement refines candidates within them. In a Lamarckian implementation, the improved solution is written back into the inherited representation. In a Baldwinian implementation, improvement affects evaluated performance without necessarily changing the genotype. Local search may consume most of the evaluation budget, making budget accounting essential when judging the value of the evolutionary component.\n\nRelated evolutionary methods differ in emphasis rather than belonging to perfectly separate categories. Genetic programming evolves programs or executable structures. Evolution strategies often emphasize real-valued mutation and adaptation of variation parameters. Differential evolution generates vector differences to construct trial solutions. Estimation-of-distribution algorithms replace explicit recombination with sampling from a learned model of promising candidates.\n\nA further extension is coevolutionary search, in which fitness depends on interactions with other evolving individuals or populations. Examples include competing strategies and cooperating components of a larger solution. Such systems can exhibit changing evaluation standards, cycles, or progress against current opponents without improvement against fixed external tests. Their dynamics therefore require different evaluation practices from optimization against an unchanging objective."
    },
    {
      "heading": "Implementation, computational cost, and parallel execution",
      "text": "The dominant cost of a genetic algorithm is often candidate evaluation rather than selection or crossover. A structural simulation, a manufacturing model, or the training of a predictive system can take far longer than manipulating its encoding. If a run evaluates approximately \\(N\\) candidates per generation for \\(G\\) generations and each evaluation costs \\(C_f\\), a useful first approximation is \\(O(GNC_f)\\), plus initialization, variation, and survivor-selection overhead. This is a cost model, not a general complexity theorem: objective cost may vary between candidates, and some implementations reuse evaluations.\n\nDuplicate detection and caching can save work when identical candidates recur. Whether caching is valid depends on the evaluator. A deterministic objective gives the same result whenever a candidate is repeated. A noisy simulation or stochastic learning procedure may produce different outcomes, so reusing a single cached observation changes the statistical interpretation of fitness. Repeated evaluation can estimate mean performance, but allocates fewer evaluations to exploring distinct solutions. Population management and evaluation policy must therefore be designed together.\n\nCandidate evaluations often support parallel computing because many can run independently. In a master–worker model, a coordinator distributes evaluation tasks and collects results. In an island model, several populations evolve with occasional migration. Island models affect search as well as execution: infrequent migration can preserve different search trajectories, whereas frequent migration can make the islands behave more like one population. Cellular algorithms restrict mating and replacement to neighborhoods, producing another form of spatially structured evolution.\n\nSynchronous execution waits for an evaluation batch to finish before advancing. Asynchronous execution incorporates results as workers complete them, reducing idle time when evaluation duration varies. However, asynchronous replacement can introduce evaluation-time bias if faster candidates have more opportunities to reproduce. A graphics processing unit can accelerate workloads with suitable regularity and batching, but irregular simulations, branching decoders, and communication overhead may limit the benefit.\n\nFor very expensive objectives, surrogate models approximate fitness using previously evaluated candidates. A surrogate can screen offspring, estimate promising regions, or decide which candidates warrant exact evaluation. Its predictions are uncertain, especially in sparsely sampled regions. If selection repeatedly favors errors in the surrogate rather than genuine performance, optimization can exploit the model’s inaccuracies. Reliable surrogate-assisted schemes therefore maintain some connection to the original evaluator and track the cost of both model construction and exact evaluation.\n\nReproducible implementation requires recording the representation, initialization distribution, random seed policy, operators, parameter values, replacement rules, and stopping conditions. Saving a final candidate without its decoder or evaluation configuration may be insufficient to reconstruct the result. Software correctness also matters: an indexing error in a permutation operator or a stale fitness value can alter apparent search performance while leaving the overall program seemingly functional."
    },
    {
      "heading": "Applications in engineering, scheduling, and machine learning",
      "text": "Genetic algorithms are used where candidate solutions can be evaluated more readily than they can be constructed optimally. In engineering design, a chromosome may specify dimensions, materials, component choices, or control parameters. A simulator computes quantities such as weight, stress, drag, efficiency, or cost. Evolutionary search can then explore alternatives without requiring analytical derivatives. Mixed discrete–continuous design is particularly compatible with this approach, although strong physical constraints and expensive simulations often require specialized encodings, repair, or surrogate models.\n\nA documented aerospace example is the evolved X-band antenna design developed for NASA’s Space Technology 5 mission, launched in 2006. Evolutionary design explored unconventional wire geometries evaluated through electromagnetic simulation under mission requirements. This case illustrates how automated search can discover useful configurations outside familiar design families. It does not establish that genetic algorithms outperform conventional electromagnetic optimization in general; its significance lies in the successful integration of representation, simulation, search, fabrication, and testing.\n\nScheduling and routing provide another major application area. A schedule may assign jobs to machines, allocate workers to shifts, or order tasks subject to precedence constraints. The chromosome can encode an operation sequence or priorities interpreted by a schedule-building procedure. Objectives may include completion time, lateness, resource utilization, or transportation cost. Because feasibility and local structure are central, successful systems commonly combine specialized crossover with local improvement rather than applying unrestricted binary operators to arbitrary encodings.\n\nIn machine learning, genetic algorithms have been used for feature selection, model configuration, rule discovery, and hyperparameter search. A binary chromosome can indicate which input variables a model uses, while another part of the representation specifies model settings. Fitness may combine predictive performance with a penalty for the number of selected features. Each candidate can require training and evaluating a model, making these searches substantially more expensive than evaluating a simple mathematical function.\n\nEvolutionary methods can also optimize the weights or architecture of an artificial neural network, a broad area known as neuroevolution. Direct weight evolution searches numerical parameters; architecture evolution searches network structures or construction rules. These methods are distinct from backpropagation, although hybrid systems can use gradients to train weights while evolution searches architectures or other configurations. Their relative usefulness depends on evaluation cost, differentiability, search dimensionality, and the task’s feedback structure.\n\nData-dependent applications introduce a specific risk: the search can overfit the evaluation procedure itself. Repeatedly choosing candidates according to one validation set can exploit its accidental characteristics. An untouched test set is needed for a final performance estimate, and procedures such as nested cross-validation can separate model selection from evaluation when data are limited. Genetic search does not remove these statistical requirements merely because it is gradient-free."
    },
    {
      "heading": "Empirical evaluation, limitations, and open problems",
      "text": "A genetic algorithm should be evaluated as a complete experimental procedure, not only through its best reported run. Random initialization and stochastic variation produce a distribution of outcomes. Independent runs, a clearly stated computational budget, and reporting of central tendency and dispersion help distinguish reliable performance from occasional success. Useful measures include final objective value, probability of reaching a target, evaluations to a target, wall-clock time, and feasibility rate. The appropriate measure depends on whether the application values solution quality, reliability, or rapid response.\n\nComparisons require consistent accounting. Equal generation counts do not imply equal effort when population sizes differ. Equal numbers of objective evaluations may still conceal major differences in local-search cost, surrogate training, or parallel hardware. Parameter tuning also consumes resources: comparing a heavily tuned genetic algorithm against an untuned baseline can confound algorithm design with experimental effort. Suitable baselines include random search, domain-specific heuristics, local search, and exact methods where the instance size permits them.\n\nProblem properties strongly affect suitability. Smooth continuous objectives with inexpensive, reliable gradients often favor gradient-based optimization. Structured combinatorial problems may admit specialized mathematical programming methods that provide strong bounds or certificates. Genetic algorithms can be useful when those assumptions fail or when multiple solution forms must be explored, but flexibility alone does not establish efficiency. Their computational complexity is not made favorable simply by avoiding derivatives or maintaining a population.\n\nNoise and nonstationarity create additional difficulties. A candidate may win selection because of a favorable random evaluation rather than superior expected performance. Repeated measurements, uncertainty-aware comparison, or independent reevaluation of finalists can reduce this effect, but all consume resources. When the objective changes over time, preserving old elites can be counterproductive unless they are reevaluated. Diversity and stored alternatives may help adaptation, yet the relevant performance measure becomes tracking a changing target rather than converging to one fixed optimum.\n\nMultiobjective evaluation poses its own measurement problems. A set with many candidates is not necessarily a better approximation to a Pareto front. Hypervolume measures the objective-space region dominated relative to a chosen reference point, making that reference and objective scaling consequential. Distance-based indicators may require a known or estimated reference front. Reporting only a visually attractive plot can obscure poor coverage, dominated solutions, or different computational budgets.\n\nContinuing research addresses how to learn useful variable interactions, adapt operators without excessive tuning, allocate evaluations under uncertainty, and exploit parallel resources without distorting selection. Another unresolved practical issue is transferability: parameters and encodings that work well on a benchmark collection may perform poorly on new instances with different structure. Analysis of failure cases, operator ablations, and evaluation on genuinely held-out problem instances provide evidence about which components exploit reusable structure and which merely fit the experimental setting."
    }
  ],
  "links": [
    "randomized-algorithm",
    "genetics",
    "mathematical-optimization",
    "heuristic",
    "pseudocode",
    "bit",
    "markov-chain",
    "no-free-lunch-theorem",
    "pareto-efficiency",
    "parallel-computing",
    "graphics-processing-unit",
    "derivative",
    "machine-learning",
    "feature-selection",
    "hyperparameter",
    "artificial-neural-network",
    "backpropagation",
    "overfitting",
    "validation-set",
    "test-set",
    "nested-cross-validation",
    "computational-complexity"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}