AI Glossary / Graphs and symbolic AI / Intermediate

Evolutionary algorithms

Definition

Evolutionary algorithms are a family of search and optimization methods inspired by variation and selection in biological evolution. They evaluate candidate solutions, create modified or recombined candidates, and use fitness-related rules to influence which candidates reproduce or survive. The aim is to find useful solutions under a defined objective, without requiring every method to use gradients.

What it means

A candidate might encode a schedule, a set of model parameters, a program, or another structured choice. Its fitness is an evaluation assigned by the problem designer. Better fitness means better performance according to that evaluation, which may differ from the full real-world objective.
The family includes genetic algorithms, evolution strategies, and other related approaches. Their representations, variation operators, and selection rules differ. The biological vocabulary is an inspiration for algorithm design; it does not imply that candidate solutions are alive or that the search has intentions.

How it works

Initialize one or more candidates, evaluate them, and repeat a cycle of selecting parents, creating offspring through variation, evaluating offspring, and deciding which candidates remain. Many familiar methods maintain a population; some evolutionary strategies use a single retained parent with offspring generated around it.
Mutation changes a candidate, while recombination combines information from parents. Not every algorithm uses both. Selection creates a preference for better evaluated candidates, and replacement determines the next population. Stop according to a stated budget or criterion. Constraints need explicit handling because variation can generate invalid solutions, and noisy fitness evaluations may require repeated measurements.

Three practical examples

• A scheduling experiment evolves valid task orderings and evaluates them with a simulator that measures delay.
• A model-development study searches over a defined set of hyperparameters using validation performance as one evaluation component.
• A design exercise evolves candidate shapes or configurations under a simulated objective and then checks promising candidates against separate requirements.

Think of it this way

Think of maintaining a workshop shelf of prototypes, testing them, and making revised prototypes from promising designs. This captures repeated evaluation and variation. The analogy is limited because algorithmic selection follows explicit rules and a numerical or structured fitness measure; it does not understand the product's purpose or notice unmeasured defects by itself.

Worked example

Use four-bit strings and define fitness as the number of ones. Suppose the population contains 1000, 1100, 0110, and 0001, with fitness values 1, 2, 2, and 1. A particular selection outcome chooses parents 1100 and 0110.
With a crossover cut after the second bit, combining the first parent's prefix 11 with the second parent's suffix 10 produces 1110, with fitness 3. The complementary child 0100 has fitness 1. Recombination therefore improved one child while producing another that is worse than either parent.
If a later mutation flips the last bit of 1110, the result is 1111 with fitness 4, the maximum possible in this toy problem. This is one possible sequence, not a prediction of a random run. A mutation could instead remove a one. The example shows how evaluation and replacement must accompany variation; variation alone does not guarantee improvement.

Illustrative case study

In a fictional planning seminar, the Juniper team evolves small timetables. Its first fitness function rewards short completion time but forgets that one resource cannot serve two tasks simultaneously. Some high-scoring candidates exploit the missing rule.
The team adds feasibility checks and compares the revised search with a simple baseline under the same evaluation budget. It records several independent runs because random choices affect the result. This invented case illustrates objective design and evaluation practice. It does not claim that an evolutionary method outperformed another optimizer or produced a deployable timetable.

Common misconceptions and mistakes

Evolutionary algorithms are not guaranteed to find a global optimum within a practical budget. A population may lose diversity and concentrate around a poor region. More generations are not automatically useful if the representation or fitness measure is wrong.
Do not assume the best individual survives unless the replacement rule preserves it. Elitism can retain strong candidates, but excessive copying may reduce exploration. Also avoid treating one successful random run as typical behavior. Report evaluation counts, constraints, and variability so comparisons reflect the actual search effort.

Business use

The family is relevant when candidate solutions can be evaluated but gradients are unavailable, inconvenient, or inappropriate. It can also provide flexible search over mixed or structured choices. The cost of evaluating each candidate is often a major practical consideration.
Before choosing a method, define the representation, objective, feasibility rules, and baseline alternatives. Reserve independent checks for promising results, especially when repeated search could overfit a validation measure. The value lies in the quality of verified candidate solutions under the real constraints, not in the biological terminology.

People also ask

Are genetic algorithms and evolutionary algorithms identical terms?

A genetic algorithm is one member of the broader evolutionary family.

Must every method use crossover?

No. Some rely primarily on mutation and selection.

Does a higher fitness score always mean a better real-world solution?

Only to the extent that the evaluation captures the relevant objective and constraints. A flawed fitness function can reward undesirable behavior.

Check your understanding

Why does the crossover example not show that recombination always improves solutions?

Answer: It produces both a child with fitness 3 and a child with fitness 1 from parents with fitness 2. The outcome depends on the parents, operator, and objective, and selection or replacement must decide what to retain.

Related terms

• Genetic algorithms
• Fitness function
• Mutation
• Selection
• Elitism
• Population
• Genetic algorithm
• evolution strategy
• crossover
• multiobjective optimization

Learn next

Learn next about representation and mutation. Encode a three-task ordering and design a mutation that always preserves a valid permutation. Compare it with independently changing positions, which can duplicate or omit tasks. Explain how the representation changes the burden placed on feasibility checks.

• Mutation

More in Graphs and symbolic AI

• Eulerian - A graph is Eulerian when it has a circuit that traverses every edge exactly once and returns to its starting vertex.
• Evidence - Evidence, in probabilistic AI, is the information on which an inference is conditioned.
• Evidence reversal - Evidence reversal is a probabilistic inference technique that restructures a model so observed evidence can guide the generation of hidden-state samples.
• Evolutionary strategies - Evolutionary strategies, commonly called evolution strategies or ES, are optimization methods that improve candidate solutions through variation...
• Evolving intelligent system - An evolving intelligent system is a learning system that can adapt its internal structure as new data arrive, as well as update numerical parameters.
• Exact cell decomposition - Exact cell decomposition is a motion-planning approach that partitions the collision-free configuration space into cells while preserving its relevant...

Sources

Sean Luke, original Essentials of Metaheuristics, population methods: https://people.cs.gmu.edu/~sean/book/metaheuristics/Essentials.pdf DEAP official evolutionary-algorithm documentation: https://deap.readthedocs.io/en/stable/api/algo.html

Previous: Evolutionary acquisition of neural topologies Next: Evolutionary computation