Can one individual be selected more than once?

Yes, in selection methods that sample with replacement. This is common in parent selection.

Does selection always use random choices?

No. Some operators select deterministically, while others use random sampling.

Is tournament selection a competition between running programs?

Not necessarily. It usually means comparing the stored fitness of a sampled subset of candidates, without a literal interaction between them.

AI Glossary / Graphs and symbolic AI / Intermediate

Selection

Definition

Selection in an evolutionary algorithm is the process of choosing individuals from a population for reproduction or survival according to a defined rule. It influences which candidate solutions contribute to the next stage of the search, often favoring those with better measured fitness.

What it means

An evolutionary algorithm maintains a population of candidate solutions. Selection decides which candidates receive opportunities to reproduce or remain in the population. Crossover and mutation are different operations: they create variation rather than simply choosing among existing individuals.

Selection is not always “keep only the best.” A probabilistic rule can give better candidates a greater chance while retaining opportunities for others. This can help preserve useful variation, although no selection rule guarantees a globally optimal result.

The term has many other uses in computing. This entry follows the source-list sense of selection in evolutionary algorithms, rather than database row selection, feature selection or choosing the next branch in a game-tree search.

How it works

Evaluate candidates using the chosen fitness definition. Make clear whether larger or smaller values are preferred and whether there are several objectives. The selection operator must interpret those values consistently.

Fitness-proportionate selection assigns probabilities from suitably positive fitness weights. Tournament selection samples a group of candidates and selects the best within that group according to the relevant ordering. Rank-based methods use relative position rather than raw score magnitude.

Distinguish parent selection from survivor selection. Choosing a candidate as a parent does not necessarily preserve that exact candidate in the next population. An elitist survival rule can explicitly retain high-quality individuals, but that is a separate design decision.

Selection pressure describes how strongly the rule favors better candidates. Strong pressure can concentrate the population quickly, while weaker pressure can retain more variety. The useful balance depends on the problem, representation, variation operators and reliability of the fitness estimates.

Three practical examples

A schedule optimizer selects parent schedules for recombination using a tournament based on constraint satisfaction and objective quality.
An evolutionary design experiment assigns reproduction probabilities from positive fitness scores and tracks whether the population loses diversity.
A genetic-programming system uses a survivor-selection rule that considers both task performance and program complexity.

Think of it this way

Think of choosing recipe candidates for another round of experimentation. Some successful recipes are tested again, while a few different candidates remain available because their ingredients may lead to useful combinations.

The analogy does not mean software candidates learn or compete consciously. Fitness is a score defined by the designer, and a recipe that scores well on a narrow test may still be unsuitable for the real task.

Worked example

Suppose four candidates A, B, C and D have positive fitness weights 8, 4, 2 and 2, with larger values preferred. Their total weight is 16.

Under fitness-proportionate parent selection, the probabilities are 8/16=0.5 for A, 4/16=0.25 for B, and 2/16=0.125 for each of C and D. If four parent slots are sampled independently with replacement, the expected selection counts are 2, 1, 0.5 and 0.5.

Those expectations are not required outcomes. A possible sample is A, A, B, D; another can omit A entirely. The probability that A is never selected in four independent draws is (1−0.5)^4=1/16.

These weights cannot simply be replaced by raw costs in a minimization problem, because larger costs would then get more selection probability. A suitable transformation or a different operator is needed, with its meaning documented.

Illustrative case study

In a fictional project, Maple Evolution minimizes delivery-plan cost. Its initial implementation feeds raw costs directly into a roulette-style selector, so expensive plans receive more reproductive opportunities.

The team notices the mismatch between objective direction and selection weights. It adopts an operator appropriate to the fitness ordering and checks a small population manually before continuing larger experiments.

The case illustrates an implementation error in selection semantics, not a general weakness of evolutionary search. It does not report real delivery savings or establish which operator is best across problems.

Common misconceptions and mistakes

Selection creates mutations. Selection chooses candidates; mutation changes them.
The best individual must always be chosen. Probabilistic selection can omit it in a particular draw.
Parent selection guarantees survival. A parent can contribute offspring without remaining itself.
Raw cost can always be used as a roulette weight. The direction and positivity requirements matter.
Stronger selection is always better. Rapid concentration can reduce diversity and trap the search.
Fitness is an objective measure of real-world worth. It reflects the chosen evaluation and can be noisy or incomplete.

Business use

When using an evolutionary optimizer, document the selection rule alongside the objective and variation operators. An unexplained “genetic algorithm” label is insufficient to reproduce the search behavior.

Monitor both solution quality and population diversity across multiple runs. A lucky seed or a noisy fitness evaluation can make one run look unusually good. Distinguish repeatable tendencies from a single best result.

Check whether candidates violating mandatory requirements can be selected and how feasibility is handled. Selection pressure should advance the actual business objective rather than reward a convenient proxy or an incorrectly directed score.

Check your understanding

A minimization problem has candidate costs 10 and 30. If those numbers are used directly as positive roulette weights, which candidate is more likely to be selected, and why is that a problem?

Answer: The cost-30 candidate receives probability 30/40=0.75. This favors the worse cost, contradicting the intended minimization objective.

Related terms

Evolutionary algorithms
Genetic algorithms
Fitness function
Mutation
Evolutionary algorithm
genetic algorithm
tournament selection
fitness-proportionate selection
crossover
selection pressure

Learn next

Create a five-candidate population with known scores. Compare a fitness-proportionate rule with a small tournament rule, explicitly stating replacement and tie handling. Explain which candidates can be chosen and whether any are guaranteed to survive.

Sources

Russell and Norvig's genetic-algorithm discussion: https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap04.pdf

DEAP selection-operator documentation and implementation, including roulette assumptions: https://deap.readthedocs.io/en/master/_modules/deap/tools/selection.html

Relevant descriptions were reviewed. The probability calculation and Maple Evolution case are original illustrations, not benchmark results.

Previous: Segmentation Next: Selection And Sampling Bias