AI Glossary / Graphs and symbolic AI / Intermediate

Elitism

Definition

Elitism is an evolutionary-search mechanism that preserves one or more of the best current candidate solutions when forming the next population. In a common genetic-algorithm version, those candidates are copied unchanged while other population members are produced through selection, crossover, and mutation. The purpose is to avoid losing a strong solution solely through replacement randomness.

What it means

An evolutionary algorithm repeatedly creates and evaluates populations of candidates. Even if one generation contains a very good candidate, the next generation might lose it when parents are sampled or offspring are altered. Elitism reserves space for selected high-performing candidates.
Best means best according to the specified fitness evaluation. It does not mean globally optimal, ethically preferable, or best on every real-world criterion. A candidate can be excellent under a flawed objective while being unsuitable for the intended application. Preserving it faithfully does not repair that mismatch.

How it works

Evaluate the current population, rank candidates according to the selection rule, and choose an elite count. Copy those candidates into the next population without the variations that would otherwise alter them. Fill the remaining positions using the algorithm's normal reproduction and replacement procedures.
Under a fixed deterministic fitness function, preserving the best candidate ensures that the best stored fitness cannot worsen simply because of generation replacement. This does not guarantee that average fitness improves or that future search finds the global optimum. Noisy fitness, changing objectives, constraints, and multiple objectives complicate what should count as elite and whether reevaluation is needed.

Three practical examples

• A genetic scheduling experiment preserves its best feasible schedule while exploring altered schedules in the remaining population.
• A parameter-search prototype retains a high-scoring configuration so random mutation cannot erase every copy in the next generation.
• A teaching simulation compares runs with no elite copies, one elite copy, and several copies to study retention and diversity.

Think of it this way

Think of keeping the best draft of a design while experimenting with new versions. The saved draft prevents accidental loss of a useful result. The analogy is limited because evolutionary populations are governed by explicit selection and replacement rules, and keeping too many copies of one draft can reduce the range of alternatives explored.

Worked example

A population has five candidates with deterministic fitness scores 12, 9, 8, 5, and 3, where higher is better. Suppose the reproduction process creates five offspring scoring 10, 8, 7, 6, and 4. Replacing the entire population would reduce the best available score from 12 to 10.
With an elite count of one, copy the score-12 candidate unchanged and fill four remaining positions with offspring. If the selected offspring score 10, 8, 7, and 6, the next population has scores 12,10,8,7,6. Its best score remains 12. Its mean is 43/5 = 8.6, compared with the original mean 37/5 = 7.4.
The mean improvement in this particular example is incidental; elitism's retention rule alone does not guarantee it. A next population of 12,1,1,1,1 would preserve the best while reducing the mean to 16/5 = 3.2. Both calculations use a fixed fitness function and unchanged elite, conditions that must be stated before claiming monotonic best-fitness retention.

Illustrative case study

At fictional Maple Designs, developers preserve several top candidates but notice that most surviving solutions become nearly identical. They compare smaller elite counts and diversity measures while keeping the same evaluation budget. They also check whether the retained candidates satisfy all required constraints.
This fictional case illustrates a retention-versus-exploration tradeoff. It does not claim that a particular elite fraction is universally appropriate or that reduced diversity necessarily prevents every later improvement.

Common misconceptions and mistakes

Elitism is not just giving a strong candidate a higher probability of becoming a parent. Probabilistic selection can still omit it, and its offspring can differ from it. Explicit unchanged survival is the key feature of the simple form described here.
Do not equate best-so-far retention with guaranteed global optimization. The search can keep a local optimum indefinitely. Copying many similar elites can consume positions that might otherwise explore different candidates. With noisy evaluations, a candidate may appear best because of a favorable random measurement; reevaluation can change its ranking. A changing objective can likewise make yesterday's elite unattractive today.

Business use

Businesses using evolutionary optimization should document the elite selection rule, count, tie handling, and feasibility requirements. Monitor both objective quality and diversity, and retain reproducible evaluation settings. If fitness is estimated from simulations or samples, distinguish uncertainty from true improvement. Compare alternative elite policies under a common computation budget. Keep an external record of useful candidates where appropriate, while recognizing that an archive and the population's survival policy serve different purposes.

People also ask

Does elitism always preserve exactly one candidate?

No. The policy can preserve one or several, depending on the algorithm.

Does it guarantee rising average fitness?

No. The best retained score can stay constant while the rest of the population worsens.

Can elite candidates still be used as parents?

Yes. Unchanged survival and participation in reproduction are separate choices; offspring can be varied while the preserved copy remains intact.

Check your understanding

The best candidate is selected as a parent but every offspring mutates, and no unchanged copy survives. Is best-candidate retention guaranteed?

Answer: No. Parent selection alone does not ensure that the original candidate remains in the next population. Its exact solution can be lost through variation.

Related terms

• Genetic algorithms
• Evolutionary algorithms
• Fitness function
• Mutation
• Genetic algorithm
• evolutionary algorithm
• survivor selection
• crossover
• diversity
• premature convergence

Learn next

Modify the worked population for a minimization problem where lower fitness is better. Specify which candidate is elite, then compare best and average values after replacement. Explain why the direction of preference must be defined before ranking candidates.

More in Graphs and symbolic AI

• Edmonds-Karp - Edmonds-Karp is an algorithm for finding a maximum flow from a source to a sink in a capacitated network.
• Effective branching factor - The effective branching factor is a summary measure that compares a search's node count with a hypothetical uniform tree of the same solution depth.
• Effective depth - Effective depth is an analytical way to describe search effort as though a search had explored a shallower tree than an uninformed baseline.
• Empirical loss - Empirical loss is the loss measured on a finite set of observed examples, usually calculated as the average of a per-example loss function.
• End-to-end learning - End-to-end learning trains a system's learnable processing stages together using an objective tied to its final output.
• English auction - An English auction is an ascending-price auction in which participants compete through increasing bids, or remain active as a displayed price rises.

Sources

Luke, Essentials of Metaheuristics, original author-hosted textbook, elitism and evolutionary algorithms: https://cs.gmu.edu/~sean/book/metaheuristics/Essentials.pdf Original author-hosted book location: https://people.cs.gmu.edu/~sean/book/metaheuristics/Essentials.pdf

Previous: Eligibility trace Next: Elliptical distribution