Evolutionary Algorithm

An evolutionary algorithm is a population-based, stochastic optimisation method inspired by biological evolution, in which a population of candidate solutions is iteratively improved through selection, recombination (crossover) and mutation guided by a fitness function. Because they require only the ability to evaluate a fitness score, evolutionary algorithms are derivative-free and well suited to non-differentiable, noisy, multimodal or black-box optimisation problems where gradient methods struggle.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:GeneticAlgorithm))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:FitnessFunction))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:SelectionOperator))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:CrossoverOperator))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:MutationOperator))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:Population))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:Neuroevolution))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:hasPart ai:QualityDiversityAlgorithm))

Dependency Relationships

SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:requires ai:FitnessFunction))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:requires ai:Sampling))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:requires ai:PopulationRepresentation))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:requires ai:FitnessLandscape))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:requires ai:PopulationDiversity))

Capability Relationships

SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:enables ai:NeuralArchitectureSearch))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:enables ai:HyperparameterOptimisation))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:enables ai:MultiObjectiveOptimisation))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:enables ai:Neuroevolution))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:enables ai:AlgorithmDesign))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:enables ai:ReinforcementLearning))

Implementation Relationships

SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:implements ai:SearchAlgorithm))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:implements ai:Metaheuristic))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:implements ai:OptimisationAlgorithm))

Reduction Relationships

SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:reducesTo ai:GeneticAlgorithm))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:reducesTo ai:EvolutionStrategy))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:reducesTo ai:GeneticProgramming))
SubClassOf(ai:EvolutionaryAlgorithm
  ObjectSomeValuesFrom(ai:reducesTo ai:DifferentialEvolution))

About

Evolutionary computation traces its origins to the independent but convergent work of three groups in the 1960s. John Holland at the University of Michigan developed genetic algorithms, inspired by the chromosome metaphor of natural heredity; his 1975 book “Adaptation in Natural and Artificial Systems” established the theoretical framework via schema theory, demonstrating that a genetic algorithm implicitly evaluates an exponential number of hyperplane cross-sections (schemata) in each generation — the so-called “implicit parallelism” that explains why genetic algorithms explore efficiently. Independently, Ingo Rechenberg and Hans-Paul Schwefel at the Technical University of Berlin developed evolution strategies (Evolutionsstrategie), using self-adaptive real-valued mutation parameters — the first adaptive mutation rate control, which remains a theoretical centrepiece of modern continuous evolutionary optimisation. Lawrence Fogel in San Diego developed evolutionary programming, which evolved finite-state machines for prediction tasks, emphasising behavioural rather than structural similarity between parent and offspring. A third convergent tradition, genetic programming — evolved computer programs rather than parameter vectors — was popularised by John Koza’s 1992 book. These threads unified under the umbrella term “evolutionary computation” in the early 1990s, forming the community around the journal IEEE Transactions on Evolutionary Computation (founded 1997) and the GECCO and PPSN conference series. The field’s intellectual charter is the No Free Lunch theorem (Wolpert and Macready, 1997), which proves that averaged over all possible objective functions, no optimisation algorithm outperforms any other — evolutionary algorithms are not universally superior but are structurally matched to the class of objectives that characterises engineering design, ecological simulation, and combinatorial planning.

The biological metaphor runs deep but is not literally accurate. Natural evolution operates on genotypes (DNA sequences) that are transcribed and translated into phenotypes (organism bodies) through development, and selection acts on phenotype fitness in a specific ecological context. Computational evolutionary algorithms compress this multi-level process: the “genome” is a data structure representing a candidate solution, “transcription” is either absent (direct encoding) or a simple mapping (indirect encoding via growth rules or neural development models), and “fitness” is an objective function value or a simulation result. This compression enables orders-of-magnitude faster evolution (millions of generations in minutes, rather than millions of years) but loses the open-ended generativity of biological evolution and the robustness of developmental buffering (canalization). The field has increasingly turned to quality-diversity algorithms — MAP-Elites, AURORA, Open-Ended Learning (OEL) — as a response to the limitation that single-objective evolutionary search converges rather than diversifies, seeking to maintain large archives of behaviorally diverse solutions rather than a single best solution.

The relationship between evolutionary algorithms and Gradient Descent defines one of the central debates in optimisation methodology. Gradient-based methods are catastrophically sample-efficient when gradients are available (learning rate times gradient norm per step), but require differentiability, unimodality (for convergence guarantees), and access to the computational graph. Evolutionary algorithms are gradient-free but dramatically less sample-efficient in high dimensions — a CMA-ES with population size proportional to the square root of the dimension still requires order-of-magnitude more objective evaluations than gradient descent for smooth, unimodal objectives. The practical consequence is a clear division of labour: gradient-based methods dominate differentiable machine learning (backpropagation for Neural Network training, policy gradient for differentiable Reinforcement Learning); evolutionary algorithms dominate black-box, discontinuous, or multi-objective settings (hardware-in-the-loop optimisation, combinatorial scheduling, Neural Architecture Search, policy search in non-differentiable environments). The emergence of Large Language Model systems as evolutionary operators in 2024–2025 has opened a third paradigm: LLM-guided evolutionary search, in which language model proposals replace random mutation by conditioning on the history of past evaluations and generating targeted candidate improvements — a form of model-based evolutionary search that dramatically improves sample efficiency on structured discrete problems.

Components / Architecture

Population representation: The encoding scheme for candidate solutions determines which genetic operators are meaningful and which crossover operations preserve building blocks. Binary encodings (original genetic algorithm substrate) support simple crossover and bitwise mutation but require a mapping to continuous problem domains. Real-valued encodings (native to evolution strategies) support Gaussian mutation and arithmetic recombination, directly preserving gradient structure in smooth sub-regions of the landscape. Tree encodings (genetic programming) support subtree crossover and point mutation, enabling evolution of programs and symbolic expressions. Neural network weight encodings (neuroevolution) can be direct (all weights as a parameter vector) or indirect (HyperNEAT encoding network patterns through a compositional pattern-producing network).

Selection mechanisms:

• Tournament selection: randomly sample k individuals, reproduce the fittest — controls selection pressure via tournament size k

• Roulette-wheel (fitness-proportionate) selection: probability of reproduction proportional to fitness value — prone to premature convergence when fit individuals dominate

• Rank-based selection: reproductive probability proportional to fitness rank rather than value — reduces dominance effects of outlier fitness values

• Elitism: always carry forward the top-k individuals unchanged — guarantees monotone improvement in the best fitness observed

• NSGA-II crowding distance: in multi-objective optimisation, prefer individuals at less dense regions of the Pareto front, maintaining front diversity

  Variation operators:

• One-point crossover: split parent chromosomes at a random locus, swap tails — disrupts long building blocks

• Uniform crossover: swap each gene position independently with probability 0.5 — position-independent mixing

• SBX (Simulated Binary Crossover): mimics single-point binary crossover in real-valued space, preserving mean of parents

• Differential mutation (DE): perturb a base vector by a scaled difference of two randomly chosen population members — DE/rand/1, DE/best/1 and variants

• Gaussian mutation (ES): add zero-mean Gaussian noise scaled by the strategy parameter σ — self-adaptive σ update enables CMA-ES

• LLM mutation (EvoLLM): use a language model’s next-token prediction to propose improved candidate solutions conditioned on the history of (solution, fitness) pairs

  Step-size adaptation (CMA-ES): The Covariance Matrix Adaptation Evolution Strategy maintains a full covariance matrix of the search distribution, updated via cumulative step-size adaptation (CSA) for isotropic step size and rank-one and rank-μ updates for directional adaptation. This enables CMA-ES to automatically recover the local gradient direction, rotate the search distribution to align with elongated fitness contours, and escape saddle points — properties that make CMA-ES the de facto gold standard for continuous black-box optimisation up to roughly 200 dimensions.

  Multi-objective frameworks:

• NSGA-II (Deb et al., 2002): non-dominated sorting with crowding distance diversity maintenance — dominant paradigm for 2-4 objectives

• NSGA-III (Deb and Jain, 2014): reference-point-based diversity for many-objective (5+ objective) problems

• MOEA/D (Zhang and Li, 2007): decomposition into scalar subproblems using weight vectors — computationally efficient, strong on structured Pareto fronts

• SMS-EMOA: hypervolume-indicator-based selection ensuring maximum hypervolume contribution

• MAP-Elites (Mouret and Clune, 2015): quality-diversity illumination algorithm that fills a feature-behaviour space with best-performing solutions per cell

  Use Cases / Major Families

  Neuroevolution and neural architecture search: One of the highest-impact contemporary applications of evolutionary algorithms is the discovery and training of neural network architectures. NEAT (NeuroEvolution of Augmenting Topologies, Stanley and Miikkulainen, 2002) simultaneously evolves network connectivity and weights, enabling discovery of minimal network architectures for specific tasks. Large-scale neuroevolution applied to deep networks (OpenAI Evolution Strategies paper, Salimans et al., 2017) demonstrated competitive performance with deep Reinforcement Learning baselines on Atari and MuJoCo benchmarks by treating the entire Neural Network weight vector as a genome and applying ES with parameter-perturbation noise — critically, this approach is easily parallelised across thousands of workers, scaling wall-clock time linearly with resources. Neural Architecture Search via evolutionary algorithms (Real et al., 2017 AmoebaNet; Real et al., 2019 Regularized Evolution) produced architectures competitive with human-designed networks on ImageNet classification while requiring no architectural a priori knowledge.

  Engineering design optimisation: Aerodynamic shape optimisation (aerofoil design, wind turbine blade geometry, automotive body drag reduction) routinely uses evolutionary algorithms because the objective — aerodynamic lift-to-drag ratio evaluated via computational fluid dynamics simulation — is expensive, non-differentiable with respect to shape parameters in the complex-geometry sense, and multimodal. Antenna design for telecommunications involves optimising complex 3D geometries for radiation pattern properties that have no closed-form gradient; NASA’s ST5 satellite used an evolutionary algorithm to design its antenna array, producing an unintuitive branching structure that outperformed human-designed alternatives. Structural optimisation in civil and mechanical engineering similarly exploits the black-box interface — structural analysis finite-element codes are treated as fitness evaluators, and the algorithm drives topology optimisation without requiring gradient-mode adjoint solvers.

  Drug discovery and molecular design: The molecular optimisation problem — maximising predicted binding affinity, bioavailability, and synthetic accessibility of drug candidates subject to Lipinski rule-of-five constraints — is naturally suited to evolutionary algorithms because molecules have complex graph-structured representations, and structure-activity relationships are typically non-differentiable with respect to discrete molecular graph edits. EvoMol (Leguy et al., 2020) applies evolutionary algorithms to molecular generation using SELFIES (Self-Referencing Embedded Strings) chemical language representations, outperforming VAE-based generative models on diversity and drug-likeness metrics. Multi-objective evolutionary algorithms (MOEA for drug discovery, Konig et al., 2024) optimise simultaneously for binding affinity, toxicity, synthetic accessibility, and novelty — objectives that are fundamentally conflicting and require Pareto front exploration.

  Reinforcement learning policy search: In settings where the Reinforcement Learning policy gradient is noisy, delayed, or difficult to estimate — sparse reward environments, long time horizons, non-differentiable physics simulators — evolutionary algorithms offer a complementary or superior approach. CMA-ES and evolution strategies have demonstrated competitive performance with PPO and SAC on complex continuous control tasks. Quality-diversity algorithms (QD-RL) maintain diverse behavioural archives, enabling discovery of multiple qualitatively different policy strategies that achieve high fitness — critical for robotics where a single brittle policy may fail in deployment. The Behaviour-Conditioned Actor (BCA) combines QD archives with reinforcement learning fine-tuning to produce sample-efficient discovery of diverse robust policies.

  Algorithm design and metaheuristic discovery: LLaMEA (Large Language Model Evolutionary Algorithm, 2025) uses an LLM as a mutation operator to automatically generate novel metaheuristic algorithms in Python, evaluating their performance on the BBOB continuous optimisation benchmark suite. The system discovered algorithms that exceed state-of-the-art optimisation algorithm performance on several benchmark functions — representing a qualitative leap from evolutionary algorithms that optimise parameters to systems that evolve the algorithms themselves. “Algorithm Discovery with LLMs: Evolutionary Search Meets Reinforcement Learning” (2025) extends this paradigm by combining LLM proposal with RL fine-tuning to discover algorithms for combinatorial problems. Evolution Strategies at the Hyperscale (2025) demonstrates that ES scales effectively to optimising frontier language model parameters, providing an alternative to gradient-based methods in settings where gradient access is restricted.

  Evolutionary model merging: Evolutionary optimisation of model merging recipes (Akiba et al., 2024) applies evolutionary search over merge configurations — the weights and layer assignments when combining pre-trained models via task arithmetic, TIES-merging, or DARE — achieving significant capability improvements over single fine-tuned models without additional training data, and producing cross-lingual and cross-modal model combinations that gradient-based training cannot easily achieve.

  Academic Context

  The theoretical foundations of evolutionary algorithms are built on three pillars: schema theory (Holland, 1975), which analysed genetic algorithm behaviour through the implicit parallelism of schema sampling; the No Free Lunch theorem (Wolpert and Macready, 1997), which established that algorithm performance is matched to problem structure; and the convergence theory of evolution strategies (Beyer, 2001; Hansen, 2011), which proved asymptotic convergence rates for CMA-ES on a broad class of smooth objectives. Empirical algorithmics for evolutionary computation matured through the COCO / BBOB benchmarking framework (Hansen et al., 2009), which provides a standardised set of 24 real-parameter optimisation benchmark functions across scalable dimensions and an automated comparison infrastructure that has enabled rigorous comparative studies of evolutionary optimisation algorithms across three decades.

  Multi-objective evolutionary computation was systematised by Kalyanmoy Deb’s NSGA-II (2002), which introduced non-dominated sorting and crowding distance as a clean two-criterion archive management method, becoming the most-cited paper in the field and the baseline against which all subsequent multi-objective algorithms are compared. Deb’s NSGA-III (2014) extended the framework to many-objective optimisation through reference-point-based diversity maintenance, enabling practical use on 5–15 objective problems that exceed the capacity of NSGA-II’s Pareto front diversity mechanism. The quality-diversity paradigm (Mouret and Clune, 2015; Lehman and Stanley, 2011 with Novelty Search) represents a fundamental departure from single-objective evolutionary search: instead of converging to a single best solution, QD algorithms fill an archive with the best solution for each cell of a user-defined behaviour space, producing illumination maps of the performance landscape.

  Neuroevolution experienced a Renaissance following the OpenAI Evolution Strategies paper (Salimans et al., 2017) and subsequent neural architecture search results from Google Brain (Real et al., 2017, 2019). The Science journal article “Neuroevolution insights into biological neural computation” (2025) demonstrated that neuroevolution produces network architectures that more closely recapitulate properties of biological neural circuits than gradient-trained networks, suggesting evolutionary approaches as tools for computational neuroscience. The MIT Press book “Neuroevolution: Harnessing Creativity in AI Model Design” (2025) synthesised the field’s trajectory from NEAT through quality-diversity neuroevolution to LLM-guided architecture search.

  The intersection of evolutionary computation with Large Language Model systems is a rapidly growing frontier. “Large Language Models as Evolution Strategies” (Lehman et al., 2024, GECCO) demonstrated LLMs as in-context crossover and mutation operators, outperforming Gaussian random mutation on discrete optimisation problems by exploiting semantic structure. LLaMEA (van Stein and Bäck, 2024–2025) used GPT-4 as an evolutionary operator for automated metaheuristic design, discovering competitive algorithms on BBOB benchmarks. “Evolution Strategies at the Hyperscale” (Lange et al., 2025) scaled evolution strategies to hundreds of millions of parameters, providing efficient derivative-free optimisation for settings where gradient access is unavailable or unsafe — a growing concern as frontier AI systems are increasingly deployed without exposing computational graphs to third-party optimisers.

  Current Landscape (2026)

  Evolutionary algorithms have entered a renaissance driven by three convergent forces: the growth of simulation-based engineering (where differentiability is impossible), the rise of LLM-guided evolutionary search (which dramatically improves sample efficiency on structured problems), and the scalability of evolution strategies to GPU-parallel clusters at scales previously requiring backpropagation infrastructure. The Population-Based Guiding (PBG) strategy for evolutionary neural architecture search (Nature Scientific Reports, 2025) demonstrated that synergising exploratory (diversity-maintaining) and exploitative (fitness-maximising) strategies within a unified evolutionary framework consistently outperforms either strategy alone, achieving competitive Neural Architecture Search performance with 40% fewer fitness evaluations than standard regularised evolution.

  The NSGA-III theoretical analysis paper (2025) provided the first near-tight expected-time bounds for NSGA-III on discrete many-objective problems, confirming that the population size requirement scales linearly with the size of the largest incomparable Pareto set — a theoretical result that validates the algorithm’s design and guides practitioners in population sizing for high-dimensional objective spaces. The Quality-Diversity community expanded through the development of efficient multi-fidelity QD algorithms that use cheap surrogate fitness approximations for the majority of evaluations and high-fidelity simulation only for archive-competitive solutions, reducing computation cost by 5-10x for expensive simulation-based fitness functions.

  In industry, evolutionary algorithms underpin AutoML pipelines at Google (Neural Architecture Search), Meta (ranking model optimisation), Amazon (logistics route optimisation), and Rolls-Royce / Airbus (aerodynamic design). The UK’s National Grid uses evolutionary algorithms for electricity dispatch optimisation under renewable variability constraints, and the NHS uses population-based optimisation for radiotherapy treatment planning where dose distribution optimisation is a multi-objective problem with non-differentiable biological effectiveness models. Evolution strategies have been deployed by Waymo and Zoox for simulation-based autonomous vehicle scenario generation — adversarially evolving driving scenarios to stress-test perception and planning systems.

  UK Context

  The UK has a strong tradition in evolutionary computation research, with Sheffield, Edinburgh, Manchester, Nottingham, and Exeter as primary centres. The University of Sheffield’s Natural Computation Group (founded by Natalio Krasnogor and including EPSRC-funded researchers in combinatorial optimisation and genome assembly) has produced foundational work on memetic algorithms — hybrid evolutionary algorithms that combine population-based search with local improvement heuristics — and has applied evolutionary algorithms to protein structure prediction, pharmaceutical formulation, and smart grid optimisation. Sheffield’s Advanced Manufacturing Research Centre (AMRC) integrates evolutionary optimisation into machining parameter selection and additive manufacturing build orientation, where the fitness function is a finite-element stress analysis simulation. Edinburgh’s Informatics has research in neuroevolution and evolutionary game theory, with particular contributions to open-ended evolution and the co-evolutionary dynamics that produce arms-race complexity escalation analogous to biological Red Queen dynamics. The University of Exeter’s EPSRC Centre for Predictive Modelling in Healthcare uses multi-objective evolutionary algorithms for patient stratification and treatment personalisation, balancing efficacy, toxicity, and cost objectives in cancer treatment protocol optimisation.

  Manchester’s Computer Science department contributes to combinatorial evolutionary algorithms for scheduling and vehicle routing, with applications to Greater Manchester’s transport authority (TfGM) for public transport timetabling under demand uncertainty. Leeds’ computational intelligence group applies evolutionary algorithms to financial portfolio optimisation and insurance risk modelling. Newcastle’s Digital Institute has used evolutionary algorithms for urban planning optimisation, balancing housing density, green space, transport access, and flood risk objectives — a genuine many-objective problem requiring Pareto front exploration rather than single-objective aggregation. The Alan Turing Institute’s Enrichment PhD programme regularly co-funds students working on evolutionary computation applied to industrial problems, reflecting the maturity of the field’s industrial relevance.

  The UK’s National Physical Laboratory has used evolution strategies for calibration of measurement instruments — a black-box optimisation problem where the objective is measurement accuracy evaluated against certified reference standards, with no available gradient. EPSRC’s Living with Machines project at the Alan Turing Institute applies evolutionary algorithms to historical newspaper digitisation quality optimisation, using fitness functions based on OCR accuracy evaluated against ground-truth transcriptions. The UK Research and Innovation (UKRI) Digital Security by Design (DSbD) programme funds evolutionary fuzzing research — using evolutionary algorithms to generate test inputs that maximize code coverage or trigger memory safety violations in security-critical software, with applications to embedded systems in critical infrastructure.

  Future Directions (2026–2030)

  LLM-guided evolutionary search at scale: The convergence of evolutionary computation with Large Language Model capabilities is the defining frontier of the field. Systems like LLaMEA and EvoLLM demonstrate that LLMs can serve as semantically informed mutation and crossover operators, conditioning proposal distributions on the history of evaluated solutions and their fitness scores. As inference costs drop and context windows expand, LLM-guided evolutionary search is expected to become the dominant paradigm for black-box optimisation over structured search spaces — scientific hypothesis generation, code synthesis, materials discovery, and drug design. The challenge is evaluation infrastructure: LLM-proposed candidates still require expensive oracle evaluation, and the sample efficiency advantage of LLM mutation must overcome inference costs to remain practical.

  Open-ended evolution for AI research automation: Quality-diversity and open-ended evolution algorithms are increasingly applied to the meta-problem of AI research itself — evolving training procedures, loss functions, optimiser configurations, and model architectures without human specification of the fitness landscape. Systems like AutoRL, EvoJAX (large-scale hardware-accelerated neuroevolution at Google), and the emerging “AI Scientist” paradigm use evolutionary algorithms to discover novel machine learning algorithms that outperform human-designed baselines, analogous to AlphaFold’s discovery of protein structures that eluded human structural biologists.

  Theoretical convergence with gradient-based optimisation: Research on gradient-estimated evolutionary strategies (natural evolution strategies, gradient-based meta-learning with evolution strategies) and the connection between CMA-ES and natural gradient descent is generating a unified optimisation theory that encompasses both gradient-based and gradient-free methods as special cases. This unification is likely to produce hybrid algorithms that seamlessly transition between gradient-based exploitation and evolutionary exploration as the fitness landscape structure changes — particularly valuable for non-stationary landscapes encountered in online Reinforcement Learning.

  Many-objective optimisation for real-world system design: As engineered systems grow in complexity (integrated energy systems, urban digital twins, personalised medicine protocols), the number of competing objectives exceeds the capacity of current NSGA-III-style algorithms (typically effective up to 10–15 objectives). Indicator-based algorithms using hypervolume and R2 indicators, decomposition-based algorithms with adaptive weight generation, and machine-learning-surrogate-assisted many-objective optimisation are active research directions aimed at scaling Pareto-optimal search to 20–50 objective settings.

  Neuromorphic and quantum evolutionary computation: Evolutionary algorithms running natively on neuromorphic hardware (Intel Loihi 2, IBM TrueNorth) offer dramatic energy efficiency advantages over von Neumann architecture implementations: the inherent parallelism of population evaluation maps naturally onto neuromorphic architectures that process populations as distributed spike patterns. Quantum evolutionary algorithms use quantum superposition to maintain exponentially many population members in hardware, with quantum interference providing implicit diversity management — speculative but actively investigated by Google Quantum AI and IBM Quantum as the gate-based quantum computing fault-tolerance threshold approaches.

  Research & Literature

  1. Holland, J.H. (1975). Adaptation in Natural and Artificial Systems. University of Michigan Press.
  2. Rechenberg, I. (1973). Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der biologischen Evolution. Frommann-Holzboog.
  3. Goldberg, D.E. (1989). Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley.
  4. Koza, J.R. (1992). Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press.
  5. Wolpert, D.H. and Macready, W.G. (1997). No Free Lunch Theorems for Optimization. IEEE Transactions on Evolutionary Computation, 1(1), 67–82.
  6. Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002). A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2), 182–197.
  7. Hansen, N. and Ostermeier, A. (2001). Completely Derandomized Self-Adaptation in Evolution Strategies. Evolutionary Computation, 9(2), 159–195.
  8. Stanley, K.O. and Miikkulainen, R. (2002). Evolving Neural Networks through Augmenting Topologies. Evolutionary Computation, 10(2), 99–127.
  9. Mouret, J.-B. and Clune, J. (2015). Illuminating Search Spaces by Mapping Elites. arXiv:1504.04909.
  10. Salimans, T., Ho, J., Chen, X., Sidor, S., and Sutskever, I. (2017). Evolution Strategies as a Scalable Alternative to Reinforcement Learning. arXiv:1703.03864.
  11. Real, E., Moore, S., Selle, A., Saxena, S., Suematsu, Y.L., Tan, J., Le, Q.V., and Ng, A.Y. (2017). Large-Scale Evolution of Image Classifiers. ICML 2017.
  12. Real, E., Aggarwal, A., Huang, Y., and Le, Q.V. (2019). Regularized Evolution for Image Classifier Architecture Search. AAAI 2019.
  13. Deb, K. and Jain, H. (2014). An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems with Box Constraints. IEEE Transactions on Evolutionary Computation, 18(4), 577–601.
  14. Zhang, Q. and Li, H. (2007). MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition. IEEE Transactions on Evolutionary Computation, 11(6), 712–731.
  15. Lehman, J., et al. (2024). Large Language Models as Evolution Strategies. GECCO 2024 Companion.
  16. van Stein, N. and Bäck, T. (2024). LLaMEA: A Large Language Model Evolutionary Algorithm for Automatically Generating Metaheuristics. arXiv:2405.20132.
  17. Lange, R.T., et al. (2025). Evolution Strategies at the Hyperscale. arXiv:2511.16652.
  18. Akiba, T., et al. (2024). Evolutionary Optimization of Model Merging Recipes. arXiv:2403.13187.
  19. Leguy, J., Browne, W., and Bengio, Y. (2020). EvoMol: a Flexible and Interpretable Evolutionary Algorithm for Unbiased de novo Molecular Generation. Journal of Cheminformatics.
  20. Beyer, H.-G. (2001). The Theory of Evolution Strategies. Springer.
  21. Hansen, N., Finck, S., Ros, R., and Auger, A. (2009). Real-Parameter Black-Box Optimization Benchmarking 2009: Noiseless Functions Definitions. INRIA Technical Report.
  22. Cully, A., Clune, J., Tarapore, D., and Mouret, J.-B. (2015). Robots that can Adapt like Animals. Nature, 521(7553), 503–507.
  23. Pugh, J.K., Soros, L.B., and Stanley, K.O. (2016). Quality Diversity: A New Frontier for Evolutionary Computation. Frontiers in Robotics and AI, 3, 40.
  24. Eiben, A.E. and Smith, J.E. (2015). Introduction to Evolutionary Computing. 2nd edition, Springer.
  25. Garcia-Sanchez, P., et al. (2025). Population-Based Guiding for Evolutionary Neural Architecture Search. Scientific Reports, 2025.
  26. Elhara, O., et al. (2025). The First Theoretical Approximation Guarantees for NSGA-III. arXiv:2504.21552.
  27. Mouret, J.-B. (2025). Neuroevolution: Harnessing Creativity in AI Model Design. MIT Press.
  28. Zoph, B. and Le, Q.V. (2017). Neural Architecture Search with Reinforcement Learning. ICLR 2017.

Provenance