Is it the same as a genetic algorithm?

Genetic algorithms are one part of evolutionary computation; the wider family includes evolution strategies and other approaches.

Does it need a neural network?

No. Candidate solutions can be many kinds of structured or numerical objects.

Why run an experiment more than once?

Random initialization and variation can produce different results. Repeated runs help reveal whether an apparent advantage is dependable.

AI Glossary / AI and machine learning / Advanced

Evolutionary computation

Definition

Evolutionary computation is a family of computational search methods inspired by variation and selection in biological evolution. Candidate solutions are changed, evaluated and selected repeatedly to find useful answers to an optimization or design problem.

What it means

Instead of directly calculating the best answer, an evolutionary method explores alternatives. A candidate might represent a delivery schedule, a machine setting, a mathematical expression or a neural network. Its fitness describes how well it meets the chosen objective.

The biological inspiration is an organizing idea, not a claim that software reproduces all of natural evolution. The approach can be useful when the search space is complicated or a reliable gradient is unavailable, but it can require many expensive evaluations.

How it works

An experiment defines a representation and creates initial candidates. Evaluation assigns fitness. Selection chooses candidates that influence the next stage. Mutation changes parts of a candidate; some methods also recombine material from multiple candidates. The process repeats until its stopping rule is reached.

Implementations differ in population management and selection. Constraints may be built into the representation, handled by repair or included in evaluation. The choice matters: an apparently excellent schedule is useless if it assigns one person to two simultaneous jobs. Independent checks should examine feasibility as well as score.

Three practical examples

1. A manufacturer can explore equipment settings in a simulator, balancing output quality against energy use. Final settings need confirmation with the real process.

2. A scheduling team can search combinations of shifts while enforcing staffing and availability rules. The representation should prevent or penalize impossible assignments.

3. A researcher can evolve symbolic expressions that predict a measured quantity. Compact expressions may be easier to inspect, but statistical validation remains necessary. These examples illustrate possible uses, not assured improvements.

Think of it this way

Picture a design workshop with many sketches. Reviewers keep promising ideas, change details and combine useful elements, then review another round. Evolutionary computation automates a defined version of that cycle.

The important difference is that the computer follows a scoring procedure. It does not independently recognize an omitted requirement such as easy maintenance or user comfort. Those requirements must enter the design and evaluation process.

Worked example

Suppose a toy search chooses three binary switches. Fitness is the number of switches set to one. The candidate 100 has score 1; 110 has score 2; 111 has score 3.

A mutation of 100 that flips its second bit yields 110, improving fitness by one. A later mutation that flips the last bit yields 111, the best possible score for this tiny problem. Another mutation could worsen it, so variation alone is not improvement. Selection and the treatment of earlier good solutions shape the outcome.

This example exposes the mechanics. Real objectives are rarely this straightforward, and interaction between settings can make locally promising changes disappointing.

Illustrative case study

Fictional furniture producer Northwood simulates cutting patterns to reduce material waste. Its first evolutionary experiment rewards low scrap but ignores the number of machine changeovers. It finds patterns that save board area while interrupting production too frequently.

The team revises the evaluation to report both waste and changeovers, then compares feasible alternatives with its existing planning process. Operators assess the shortlisted plans before a trial. The lesson is that search quality depends on what the evaluation recognizes, not simply on how many generations the software runs.

Common misconceptions and mistakes

An evolutionary algorithm does not guarantee a global optimum within a practical budget. A plateau can reflect premature convergence, insufficient variation or an unhelpful representation. More generations do not necessarily fix those issues.

The method also does not make data quality irrelevant. If fitness comes from a biased model, the search may exploit that model's errors. Avoid comparing algorithms using unequal evaluation budgets, reporting only the best random run or forgetting constraints. The word evolutionary should not be taken as evidence that a method can adapt safely after deployment.

Business use

Evolutionary computation can support difficult configuration and design searches where alternatives can be evaluated economically. Before funding a large experiment, estimate evaluation cost and compare simpler methods such as random search, established optimization tools or domain heuristics.

Track the objective, constraint violations, variability across runs and independent performance of the selected answer. If simulation dominates the project cost, improving the simulator or using carefully validated approximations may matter more than changing the evolutionary operator. Retain the decisions needed to reproduce a promising result.

Check your understanding

A search finds a delivery plan with the lowest distance but some deliveries miss their deadlines. Has it solved the business problem?

Answer: Only if deadlines were irrelevant, which the scenario suggests they were not. The representation or evaluation omitted an important requirement. Optimize feasible plans and assess the actual operational objective.

Related terms

Genetic algorithms
Mutation
Fitness function
Population
Multi-objective optimization
Genetic algorithm
evolution strategy
crossover
simulation optimization
random search

Learn next

Build a small candidate representation for a familiar planning problem. Identify valid changes, forbidden configurations and the information needed to score an answer. Compare a simple evolutionary search with a baseline using the same evaluation budget.

Sources

DEAP project, official overview and tutorial of evolutionary-computation components: https://deap.readthedocs.io/en/master/overview.html. This source illustrates representations, fitness, variation and selection. The switch example, production scenario and business evaluation guidance are original teaching material.

Previous: Evolutionary algorithms Next: Evolutionary data mining