Is this the same as ordinary evolutionary computation?

It uses evolutionary search, but human evaluation supplies at least part of the fitness feedback.

Must users assign numerical scores?

No. Pairwise choices, rankings and other interfaces can provide feedback.

Does it learn permanent preferences?

Not necessarily. Preferences may be local to the task or session and can change as the user explores.

AI Glossary / AI and machine learning / Intermediate

Interactive evolutionary computation

Definition

Interactive evolutionary computation is an optimization approach in which people evaluate candidate solutions and an evolutionary algorithm uses that feedback to generate further candidates. It is useful when preferences are difficult to express as a complete numerical formula.

What it means

Some design goals are easy to calculate, such as weight or cost. Others involve how a sound feels, whether a shape is appealing or which arrangement a person prefers. Interactive evolutionary computation places human judgment inside the search loop.

The computer can explore variations while the person supplies evaluations. Feedback may take the form of scores, rankings or choices between alternatives. The method does not require the person to explain every reason for a preference, although explanations can still be useful.

Human feedback is a limited resource. People become tired, change their minds and judge candidates relative to what they have just seen. A successful system therefore needs both a suitable evolutionary search and an interface that makes evaluation manageable. It is not simply a standard algorithm with an unlimited human scoring function attached.

How it works

A candidate is represented by parameters or a structure that the algorithm can modify. An initial population is presented to the user. The user evaluates the candidates, and selection favors those receiving better feedback.

Variation operators, such as mutation and recombination, create a new population. The cycle repeats until a stopping condition is reached or the user is satisfied. Automatic constraints can exclude infeasible candidates before they consume the user's attention.

Design choices include how many alternatives to show, how feedback is collected and how diversity is maintained. Some systems use a learned approximation of user preferences to reduce the number of direct evaluations, but that introduces another source of error. The final candidate should be checked against objective requirements as well as preference, especially when attractive output can conceal a functional weakness.

Three practical examples

1. A sound designer selects preferred synthesizer tones from small groups of generated alternatives, guiding later variations.

2. A product designer explores handle shapes while automatic checks exclude designs outside manufacturing limits. Human judgment addresses appearance and perceived comfort during early exploration.

3. A visual-layout tool lets a user rank alternative arrangements, then evolves new arrangements that combine and vary the preferred parameters.

Think of it this way

Think of selectively breeding a garden of design ideas: keep promising varieties and explore their variations. The limit is that designs do not inherit biological traits automatically. The representation and variation rules determine what can change, and human preferences may shift as new possibilities appear.

Worked example

Suppose an illustrative interface shows 8 candidates per round for 10 rounds. If each candidate takes an average of 6 seconds to assess, direct evaluation takes 8 × 10 × 6 = 480 seconds, or 8 minutes, before allowing for comparison, breaks and interaction.

Showing 20 candidates for the same number of rounds would require 1,200 seconds, or 20 minutes, at that rate. More candidates may explore more possibilities but can also increase fatigue and inconsistent scoring.

The calculation does not identify an optimal population size. It shows why user effort belongs in the design objective rather than being treated as a free computational resource.

Illustrative case study

A fictional audio-studio team builds a tool for exploring notification sounds. Its first interface asks users to rate many similar clips, and later ratings become hurried. Some users also change their standard after hearing a particularly pleasing sound.

The team reduces each comparison set, allows breaks and preserves a few previously preferred examples as reference points. It separately checks loudness and duration requirements before presenting candidates.

The case illustrates that the quality of feedback depends on the interaction. Improving the search algorithm alone would not address fatigue, comparison difficulty or inconsistent presentation conditions.

Common misconceptions and mistakes

Human evaluation makes the result objectively best. It reflects particular people, conditions and preferences.
More evaluation rounds always help. Fatigue and diminishing returns can reduce the value of additional feedback.
Preference can replace functional testing. A preferred design may still fail manufacturing, accessibility or performance requirements.

Business use

Interactive evolutionary computation can support exploratory design where people can recognize attractive alternatives more easily than write a scoring formula. Define which aspects are subjective and which can be checked automatically.

Budget user time, preserve promising alternatives and let people stop or revisit choices. Test whether the process helps users reach useful designs compared with simpler exploration tools. If several stakeholders are involved, make disagreement visible rather than pretending one aggregate score captures every requirement. The method is a structured collaboration between search and judgment.

Check your understanding

Why should an interactive evolutionary system reject obviously infeasible candidates automatically before asking a person to score them?

Answer: It preserves limited human attention for meaningful choices and prevents preference feedback from being wasted on unusable alternatives.

Related terms

Learn next

Read evolutionary computation and preference learning. Sketch a small design experiment and identify the candidate representation, automatic constraints, feedback method and stopping condition.

Sources

Takagi, Interactive Evolutionary Computation: Fusion of the Capabilities of EC Optimization and Human Evaluation (2001), author institution copy: https://catalog.lib.kyushu-u.ac.jp/opac_download_md/1670053/ProcIEEE_3.pdf. Takagi and Pallez, Paired Comparisons-based Interactive Differential Evolution: https://arxiv.org/abs/0909.2091. The timing calculation and studio scenario are invented.

Previous: Interactive activation and competition networks Next: Interactive machine translation