Is recombination the same as mutation?

No. Recombination combines information from parents; mutation changes a candidate through a separate variation operation.

Must a child be better than its parents?

No. The worked example produces both a higher-scoring and a lower-scoring child from the same pair.

Why can ordinary crossover fail on a route permutation?

Segment exchange can repeat some items and omit others, violating the requirement that each item appear exactly once.

AI Glossary / Graphs and symbolic AI / Intermediate

Recombination

Definition

Recombination is an evolutionary-computation operation that creates candidate solutions by combining information from parent candidates. In genetic algorithms it is commonly called crossover. Its purpose is to explore combinations of existing solution components, after which the resulting offspring must be evaluated. Recombination does not guarantee improvement and must respect the representation used for the problem.

What it means

A candidate might be encoded as a bit string, a list of parameters, a permutation, or a program tree. Recombination describes how material from parents is combined to create another candidate in that space.

The idea is inspired by biological inheritance, but the computational operation is deliberately designed. There is no general rule that half of one good solution plus half of another produces something better. Components can interact, and a combination that looks promising can be invalid or perform poorly. The objective and constraints determine what counts as a useful offspring.

How it works

Select parent candidates and apply a recombination operator suited to their encoding. A simple one-point crossover cuts two sequences at a selected position and exchanges the portions after that cut. Other operators exchange several positions, blend numerical values, or replace compatible subtrees.

Evaluate the offspring under the actual fitness function and constraint checks. A separate selection process decides which candidates remain in the population. Mutation, when used, introduces changes through a different operation rather than merely combining parental material.

Implementation details matter. Some libraries modify candidate objects directly, so preserving parents may require copying them first. Cached fitness values must also be reconsidered after a candidate changes. Otherwise, a newly recombined candidate can accidentally retain an evaluation belonging to its old contents.

Three practical examples

Configuration search: a candidate encodes several software settings, and recombination mixes settings from two previously evaluated configurations.
Feature selection: a bit string represents which input features are included, and crossover produces a new combination for evaluation.
Program evolution: a genetic-programming system exchanges compatible expression subtrees, then checks the resulting program's behavior and fitness.

Think of it this way

Think of creating a new menu by combining sections from two existing menus. You might discover a useful combination, but the ingredients, preparation capacity, and overall balance still need checking. The analogy highlights recombination rather than guaranteed improvement. In an algorithm, the allowed edits and evaluation criteria must be explicit instead of relying on a person's intuition about what fits together.

Worked example

Take two six-bit parent strings, P = 111000 and Q = 000111. Define a toy fitness function as the number of ones. Both parents have fitness three.

Cut both strings after the third bit. Combine P's first part, 111, with Q's last part, 111, to make offspring 111111 with fitness six. The complementary offspring combines 000 with 000, giving 000000 with fitness zero.

The same crossover produces one better candidate and one worse candidate under this particular objective. Recombination itself did not know which would be better; evaluation establishes the scores.

Now change the problem so that exactly three bits must be one. Both offspring become infeasible even though the parents satisfy the requirement. A high count of ones is no longer sufficient for acceptance. Constraint handling, repair, or a representation-aware operator would be needed.

A permutation gives another warning. Parents [1, 2, 3, 4] and [3, 4, 1, 2], cut after two positions, can produce [1, 2, 1, 2] using naive segment exchange. This repeats items and omits others, so it is not a valid permutation. An operator designed for permutations addresses a different requirement from ordinary bit-string crossover.

Illustrative case study

Fictional case: Ashfield Search explores combinations of optional inputs for a prediction task. Its team encodes inclusion choices as bits and evaluates recombined candidates on a defined development protocol.

Reviewers discover that some combinations exceed a feature-cost limit. They add explicit feasibility checking and keep those violations separate from predictive scores. They also preserve the original parent candidates and ensure offspring are evaluated after changes.

The case illustrates representation, constraints, and evaluation. It does not claim a measured model improvement, a completed search, or evidence that evolutionary search is preferable to another method for that task.

Common misconceptions and mistakes

Do not assume offspring inherit the parents' fitness. Fitness belongs to the complete candidate under a particular evaluation, not to its label or ancestry.

Do not use the same crossover blindly for bit strings, permutations, and program trees. Their validity requirements differ.

Avoid describing recombination as guaranteed innovation. Some combinations are duplicates, some are invalid, and some destroy useful interactions. Also distinguish recombination from selection: mixing candidates creates possibilities, while selection determines which possibilities continue. If all parents are nearly identical, recombination may introduce little diversity unless other mechanisms supply new variation.

Business use

For business optimization, recombination is one component of an experiment, not a justification for accepting its output. Teams should state how solutions are encoded, which combinations are permitted, and how performance is measured without leaking evaluation information.

In the feature-selection example, a candidate with an attractive predictive score still fails if its required inputs exceed the permitted cost. Report feasibility and quality separately. Keep enough experiment history to reproduce a candidate's ancestry, transformation, and evaluation when investigating unexpected results.

Check your understanding

Under the exactly-three-ones constraint, should 111111 be accepted because its count-of-ones score is six?

Answer: No. It violates the stated feasibility requirement. Its score under another objective does not override that constraint.

Related terms

Learn next

apply one-point crossover at different positions in the worked strings and score each offspring. Then repeat with a permutation and identify invalid results. Study representation-specific operators before introducing recombination into a constrained search problem.

Sources

Primary source: DEAP, Evolutionary Tools, crossover descriptions and operator responsibilities: https://deap.readthedocs.io/en/master/api/tools.html Relevant official documentation was read. This entry uses the evolutionary-computation sense of the broad term; the supplied category is retained. The strings, permutation, and Ashfield Search case are original teaching examples.

Previous: Reciprocity Next: Recommender System