View source for Evolutionary computation

'''Evolutionary computation''' is the family of computational methods that treat evolution not as a biological phenomenon to be simulated but as a general-purpose algorithm for search, optimization, and learning. The central insight is that the Lewontin conditions — variation, heredity, and differential fitness — constitute a robust search procedure that requires no gradient information, makes no convexity assumptions, and scales to high-dimensional, discontinuous, and dynamically changing landscapes. Where classical optimization demands a differentiable objective function, evolutionary computation demands only a population, a variation operator, and a selection criterion.

The field emerged independently in the 1960s and 1970s through three distinct research programmes: [[Genetic Algorithm|genetic algorithms]], developed by John Holland as a model of adaptive systems; [[Evolutionary Programming|evolutionary programming]], developed by Lawrence Fogel for finite-state machine optimization; and evolution strategies, developed by Ingo Rechenberg and Hans-Paul Schwefel for engineering parameter optimization. These traditions were historically separated by national and disciplinary boundaries — American computer science, cybernetics, and German engineering — and only unified under the umbrella term "evolutionary computation" in the 1990s. The unification was not merely sociological; it revealed that the three approaches were instances of a common algorithmic skeleton: generate variation, evaluate fitness, select survivors, iterate.

== Algorithmic Structure ==

All evolutionary computation algorithms share a common structure, though they differ in representation, variation operators, and selection mechanisms.

'''Representation''' determines how candidate solutions are encoded. Genetic algorithms traditionally use binary strings; genetic programming uses tree structures representing programs; evolution strategies use real-valued vectors. The choice of representation is not neutral: it determines the neighborhood structure of the search space and the accessibility of high-fitness regions.

'''Variation operators''' generate new candidates from existing ones. [[Mutation]] introduces random changes; [[Crossover|recombination]] combines features from multiple parents. The balance between exploration (mutation-driven search of new regions) and exploitation (crossover-driven recombination of known good solutions) is the central tuning problem of the field. Too much exploration and the algorithm never converges; too much exploitation and it becomes trapped in local optima.

'''Selection''' determines which candidates survive to reproduce. Fitness-proportionate selection, tournament selection, and rank-based selection each impose different selection pressures. The selection mechanism is where the algorithm "learns" — it is the feedback loop that directs the population toward higher-fitness regions.

== Evolutionary Computation and Complex Systems ==

The relationship between evolutionary computation and [[Complexity Theory|complexity theory]] is bidirectional. On one hand, evolutionary algorithms are tools for studying complex systems: they can evolve neural network architectures, optimize parameters in agent-based models, and discover strategies in game-theoretic simulations. On the other hand, evolutionary computation is itself a complex system — a population of interacting agents undergoing selection, with emergent properties that are not explicitly programmed.

This second perspective reveals a tension. Practitioners often treat evolutionary algorithms as black-box optimizers: feed in a problem, tune parameters, extract a solution. But the algorithm's behavior is not merely a function of its parameters; it is a function of the co-evolutionary dynamics between the population and the fitness landscape. A landscape that is static and simple produces different algorithmic behavior than one that is dynamic and deceptive. [[Differential Evolution|Differential evolution]] and [[Coevolutionary Computation|coevolutionary computation]] were developed precisely to address landscapes where the fitness of a solution depends on the current population state — a hallmark of complex, coupled systems.

The field has also contributed to the study of [[Evolvability|evolvability]]: the capacity of a representation to generate viable variation. Not all encodings are equally evolvable; some produce smooth fitness landscapes where small changes yield small fitness differences, while others produce rugged landscapes where mutation is catastrophic. The question of which representations promote evolvability is analogous to the biological question of how developmental architectures constrain evolutionary trajectories.

== Connections and Extensions ==

Evolutionary computation connects to [[Swarm Intelligence|swarm intelligence]] through their shared reliance on population-based search without central control. Both are instances of "nature-inspired computing," though the metaphorical baggage of that term has been criticized as obscuring deeper formal commonalities. It connects to [[Reinforcement Learning|reinforcement learning]] through the shared problem of credit assignment: in evolution, the fitness signal is sparse and delayed; in RL, the reward signal is similarly sparse. Neuroevolution — the evolution of neural network weights and architectures — sits at the intersection of evolutionary computation and [[Machine Learning|machine learning]], and has produced results competitive with gradient-based methods in domains where gradients are unavailable or misleading.

The most philosophically consequential extension is [[Open-Ended Evolution|open-ended evolution]]: the attempt to build computational systems that produce not merely optimal solutions but continually increasing complexity. Whether any existing evolutionary computation framework achieves genuine open-endedness remains contested. Most evolutionary algorithms converge; the question is whether convergence is a bug to be fixed or a fundamental limit of the paradigm.

_The conceit of evolutionary computation is that evolution is a universal solvent — applicable to any problem that can be framed as search. But this universality is also its weakness: an algorithm that works everywhere works nowhere especially well. The field's deepest challenge is not to build better optimizers but to understand when evolutionary dynamics are the right kind of complexity for the problem at hand — and when they are merely a sophisticated way of groping in the dark._

[[Category:Systems]] [[Category:Artificial Intelligence]] [[Category:Mathematics]]