Evolutionary Computation

Evolutionary computation is the branch of artificial intelligence that borrows the logic of biological evolution—variation, selection, inheritance—and turns it into a family of algorithms for search, optimization, and design.

13 frameworks, from 1962 to today · Artificial Intelligence · Computer Science

still a working view ended · each leads to its entry below

Overview

Evolutionary computation is the branch of artificial intelligence that borrows the logic of biological evolution—variation, selection, inheritance—and turns it into a family of algorithms for search, optimization, and design1. The premise is deceptively simple: maintain a population of candidate solutions, let the better ones reproduce with small random changes, and repeat1. Over many generations, good solutions accumulate. What is at stake is a different model of machine intelligence from the dominant one. Where most of AI learns by following gradients downhill or by encoding expert rules, evolutionary computation explores by generating diversity and letting selection do the work. It is the field that asks whether intelligence can emerge from blind trial and error at scale—and, in doing so, it has produced practical tools for problems that resist calculus, from antenna design to protein folding.

Precursors: Darwinian Ideas Before Computers

The intellectual seed is Darwin's. By the late nineteenth century, natural selection was understood as an algorithm-like process: heritable variation plus differential reproduction yields adaptation without a designer. Early twentieth-century mathematicians and engineers occasionally gestured at simulating this. But the decisive move came with the stored-program computer, which made populations of thousands of trial solutions cheap to evaluate. The 1940s and 1950s saw scattered experiments—Alan Turing's 1950 Mind paper speculated about "genetical or evolutionary search" as a route to machine intelligence—but no sustained research program. The field's real founding came in the 1960s, and it came from several directions at once, largely independently.

Three Foundings in the 1960s: Strategies, Programming, and Algorithms

Three distinct traditions emerged in the 1960s, each with its own vocabulary and its own view of what evolution was for.

Evolution strategies (ES) grew out of engineering optimization in Germany1. In the early 1960s, students at the Technical University of Berlin—working with Ingo Rechenberg and later Hans-Paul Schwefel—faced problems like optimizing the shape of a bent pipe to minimize drag, where the objective function was a physical experiment with no usable gradient. Their idea was to perturb a candidate solution with random changes, keep the change if it improved the result, and adjust the size of the perturbations over time. This "one individual, mutate, select" loop was later generalized to populations and to recombination. The key insight was self-adaptation: the mutation step size itself could be encoded in the individual and evolve alongside the solution, so the algorithm learned how boldly to search.

Evolutionary programming (EP) came from Lawrence Fogel in the United States, starting around 1962–19641. Fogel, working on the problem of predicting sequences (such as symbols from a finite-state machine), proposed evolving populations of finite-state machines by random mutation and selection against a prediction error1. EP's emphasis was on behavior: the population was a set of candidate programs, and fitness was how well each predicted the future. It was a direct, if early, statement that intelligence could be produced by simulated evolution of programs.

Genetic algorithms (GA) were John Holland's contribution, developed at the University of Michigan through the 1960s and crystallized in his 1975 book Adaptation in Natural and Artificial Systems. Holland's framing was the most theoretically ambitious. He encoded solutions as strings of bits (chromosomes), applied crossover (recombining parts of two parents) and mutation, and selected parents by fitness1. His central claim was the schema theorem: short, well-performing patterns of bits ("schemata") would be propagated and combined at an exponential rate, giving the GA a built-in, massively parallel search over building blocks2. Whether that theorem actually justified the algorithm's power became one of the field's longest-running debates, but Holland's framing—crossover as the engine, bit strings as the representation—defined the popular image of evolutionary computation for decades.

These three traditions barely knew of each other for years. They used different representations, different operators, and different justifications, and their communities met only gradually.

Genetic Programming and the Rise of Evolved Programs

If EP evolved finite-state machines, the next step was to evolve arbitrary computer programs. John Koza, at Stanford, developed genetic programming (GP) in the late 1980s and early 1990s, publishing Genetic Programming: On the Programming of Computers by Means of Natural Selection in 1992. GP represents individuals as tree-structured expressions—functions at the internal nodes, variables and constants at the leaves—and evolves them by crossover (swapping subtrees) and mutation1. Koza's demonstrations were striking: GP rediscovered patented inventions, produced controllers, and generated symbolic regression formulas. GP made the field's promise concrete—evolution as an automatic programmer—and it remains the branch most associated with "evolving solutions you did not design."

The 1980s and 1990s: Consolidation, Competition, and the First International Community

Through the 1980s the three traditions matured separately, each with its own conferences and journals. The turning point was their convergence. A series of workshops and the first International Conference on Genetic Algorithms (Pittsburgh, 1985) and later the Parallel Problem Solving from Nature (PPSN) conferences brought the communities together. By the early 1990s, "evolutionary computation" was the umbrella term, and researchers began comparing operators across traditions.

This period also produced the field's most important negative result and its most important theoretical clarification. In 1997, David Wolpert and William Macready published the No Free Lunch (NFL) theorems. In plain terms: averaged over all possible problems, every search algorithm performs identically3. No algorithm—evolutionary or otherwise—is universally superior; a method tuned to one class of problems must pay for that tuning on another. For evolutionary computation this was double-edged. It undercut grand claims that evolution was a general-purpose optimizer, but it also legitimized the field's practical stance: design algorithms for the structure of the problems you actually face. NFL reframed the research question from "which algorithm is best?" to "which algorithm matches which problem structure?"

The same decade saw the field's methods diversify. Memetic algorithms, named by Pablo Moscato in 1989, combined evolutionary search with local improvement—each individual was "polished" by a local optimizer before selection, blending global exploration with local exploitation. Multi-objective evolutionary optimization matured with the introduction of Pareto-based selection: rather than a single fitness score, algorithms like NSGA (Nondominated Sorting Genetic Algorithm, Srinivas and Deb, 1994) and later NSGA-II (Deb and colleagues, 2002) maintained populations spread across trade-offs between conflicting objectives, returning a whole front of good compromises rather than one answer. Estimation of distribution algorithms (EDAs), developed from the mid-1990s (notably by Mühlenbein and Paass), replaced crossover and mutation with a probabilistic model: the algorithm estimated the distribution of good solutions and sampled new ones from it, attacking the representation problem directly.

Swarm Intelligence: Evolution's Social Cousin

A parallel strand drew on a different biological metaphor: collective behavior. Swarm intelligence studies algorithms inspired by the emergent problem-solving of groups—ant colonies, bird flocks, fish schools4. Marco Dorigo's ant colony optimization (early 1990s) modeled how ants deposit and follow pheromone trails to find short paths, and applied it to combinatorial problems like the traveling salesman. James Kennedy and Russell Eberhart's particle swarm optimization (1995) modeled a flock of candidate solutions moving through the search space, each pulled toward its own best position and the swarm's best. Swarm methods are not strictly evolutionary—there is no reproduction or selection in the Darwinian sense—but they share the field's core logic of population-based search with decentralized feedback, and they are usually taught alongside it.

Neuroevolution: Evolving the Networks Themselves

One of the field's most consequential branches is neuroevolution: using evolutionary methods to design artificial neural networks—their weights, their architectures, or both5. Early work in the late 1980s and 1990s evolved network weights where gradient methods were awkward. The approach gained prominence with NEAT (NeuroEvolution of Augmenting Topologies), introduced by Kenneth Stanley and Risto Miikkulainen in 2002, which evolved network structure incrementally while protecting new structural innovations from being outcompeted prematurely. Neuroevolution's modern relevance is direct: as deep learning grew, evolutionary methods became a practical way to search over architectures and hyperparameters, and the two fields increasingly borrow from each other.

Differential Evolution and the Engineering Turn

In 1995–1997, Rainer Storn and Kenneth Price introduced differential evolution (DE), a compact and remarkably effective algorithm for continuous optimization. DE creates a mutant vector by adding a scaled difference between two population members to a third, then crosses it with a target individual and keeps the better. It has few parameters, is simple to implement, and performed strongly in the optimization competitions that became a fixture of the field from the 1990s onward. DE's rise marked a broader shift: evolutionary computation increasingly proved itself not through biological analogy but through benchmark performance and real engineering deployment.

The 2000s to the Present: Benchmarks, Hybrids, and the Deep Learning Era

Since the 2000s the field has been defined less by new founding metaphors than by rigor, hybridization, and scale. The Congress on Evolutionary Computation and the Genetic and Evolutionary Computation Conference became the main venues; standardized benchmark suites and competitions (such as the Black-Box Optimization Benchmarking workshops) made comparisons reproducible. The theoretical program matured: runtime analysis of evolutionary algorithms on well-defined problems gave precise results about how long algorithms take and why, and the NFL framework was extended and refined.

Practically, evolutionary methods found niches where gradients fail or where the search space is discrete, noisy, or multi-objective. They have been used to design antennas (NASA's evolved antenna work in the mid-2000s), to optimize aerodynamic shapes, to tune machine-learning pipelines, and to search chemical and materials design spaces. In the deep-learning era, evolutionary methods are used for neural architecture search, for reinforcement-learning policy optimization (notably the Evolution Strategies as a Scalable Alternative to Reinforcement Learning work by OpenAI researchers in 2017), and for generating diverse behaviors in open-ended simulated environments.

The questions still open are the field's oldest ones, sharpened. How much of evolution's power comes from selection, how much from the representation, and how much from the operators? When does a population-based search beat a well-tuned gradient method, and can that boundary be predicted rather than discovered by trial? And the deepest one: whether open-ended, continually innovating search—evolution's most distinctive feature in nature—can be reproduced in silico, rather than merely used as an optimizer for problems we have already posed.

References

  1. Evolutionary computation (introduction, History), Wikipedia
  2. Holland's schema theorem, Wikipedia
  3. No free lunch in search and optimization, Wikipedia
  4. Swarm intelligence, Wikipedia
  5. Neuroevolution, Wikipedia

The frameworks of Evolutionary Computation

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1962-1995
  2. 1963-Present
  3. 1975-2000
  4. 1989-Present
  5. 1989-Present
  6. 1991-Present
  7. 1992-Present
  8. 1994-Present
  9. 1995-Present
  10. 1996-Present
  11. 1997-Present
  12. 1998-Present
  13. 2011-Present
A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Evolutionary Computation is a planet in Nousfarer

Its city has a tower for each of these 13 frameworks. Land, walk in, and climb one concept at a time.

Land on Evolutionary Computation

Related fields

More in Artificial Intelligence

See where Evolutionary Computation sits on the Map of Human Knowledge