Must the surrogate be a neural network?

No. Many regression models or simplified physical models can serve as approximations, depending on the problem.

Can it replace every expensive evaluation?

Sometimes only approximation data is available, but strong claims about final quality then require appropriate limitations. Where authoritative evaluation is possible, retaining checks is valuable.

Is fitness approximation identical to Bayesian optimization?

No. Bayesian optimization is a specific model-based optimization approach; fitness approximation is a broader idea used in evolutionary search and other settings.

AI Glossary / AI and machine learning / Advanced

Fitness approximation

Definition

Fitness approximation estimates the quality of candidate solutions using a cheaper substitute for an expensive or unavailable evaluation. In evolutionary computation, that substitute may be a simplified simulation, a learned surrogate model or an estimate based on previously evaluated candidates.

What it means

An optimization algorithm may need to assess thousands of alternatives. If each assessment requires a long simulation, a physical experiment or a human judgment, evaluation can dominate the cost of the search.

An approximation helps decide where to spend that limited evaluation budget. Its predicted score is not the same as a measured result. The central practical question is how to use inexpensive guidance without allowing the optimizer to exploit errors in the approximation and select a disappointing final answer.

How it works

A team gathers evaluated examples or constructs a lower-cost model of the process. The approximation predicts fitness for additional candidates. Selected candidates receive a more authoritative evaluation, and those results may update the surrogate.

The policy deciding which predictions to trust and which candidates to evaluate is often called model management. It can balance predicted performance, uncertainty and coverage of the search space. Approximations do not have to be neural networks, and their accuracy may vary by region. A model that performs well on average can still rank the most promising candidates incorrectly.

Three practical examples

1. An engineering search can use a learned model to screen component designs before running a costly high-resolution simulation on a shortlist.

2. A manufacturing experiment can estimate performance between tested settings, then choose additional physical trials to check promising or uncertain regions.

3. An interactive design tool can approximate human preference scores to reduce repetitive ratings, while periodically asking for direct judgments. The approximation should not be treated as a complete representation of the person's preferences.

Think of it this way

Imagine using a quick sketch to decide which building designs deserve a detailed model. The sketch saves effort, but you would not assume every structural feature is correct because it looks attractive on paper.

Similarly, approximate fitness is guidance for allocating evaluation effort. It should remain connected to evidence from the more authoritative process, particularly when the search moves beyond familiar examples.

Worked example

Suppose a high-fidelity simulation takes thirty minutes per candidate. Evaluating 200 candidates directly would require 6,000 minutes, or 100 hours of serial simulation time.

In a hypothetical assisted search, forty initial simulations provide training data and twenty further simulations check selected candidates. Those sixty simulations require 1,800 minutes, or thirty hours. If fitting and querying the surrogate take two additional hours, the total is thirty-two hours, saving sixty-eight hours compared with the direct calculation.

This arithmetic assumes the stated costs and ignores parallelism. It does not prove that both searches find equally good answers. Any time saving must be weighed against validated final-solution quality and the risk of missed alternatives.

Illustrative case study

Fictional pump-design firm FlowCraft uses a surrogate to rank simulated geometries. The search proposes an unusual geometry with an excellent predicted score. A high-fidelity check shows that the prediction was overly optimistic because the design lies outside the well-sampled region.

The team adds that evaluation to its dataset and adjusts how it selects further trials. It keeps final rankings based on verified simulations for shortlisted designs. The experience demonstrates how optimization can expose weaknesses in an approximation, rather than proving that a good average prediction error makes every proposed optimum trustworthy.

Common misconceptions and mistakes

Fitness approximation does not make a computationally hard search exact. It changes how candidates are evaluated and can introduce a second source of error beyond the search itself.

Do not validate a surrogate only where training examples are dense, then trust extrapolation elsewhere. Also distinguish predicting numerical scores from correctly ranking candidates: the latter may matter more for selection. A high-fidelity simulation is itself a model, so it may still need physical validation. Finally, compare total project effort, including data collection and model maintenance, rather than counting only fast surrogate queries.

Business use

The approach is useful when evaluation is expensive enough that a carefully managed approximation can save meaningful time or cost. It is less attractive when direct evaluations are cheap or the surrogate is costly to maintain.

A business case should state the authoritative evaluation, the approximation's role and the policy for checking final candidates. Track simulation or experiment budget, ranking errors and verified solution quality. Keep predicted results visibly separate from confirmed results. This helps stakeholders understand whether an attractive score represents evidence or a proposal for further testing.

Check your understanding

A surrogate predicts its best-ever score for a candidate far outside the training region. Is that sufficient to accept the candidate?

Answer: No. The apparent advantage may be extrapolation error. Consider uncertainty and obtain a more authoritative evaluation before treating the predicted result as established performance.

Related terms

Surrogate model
Evolutionary computation
Active learning
Bayesian optimization
model management
multi-fidelity modeling
simulation optimization
extrapolation

Learn next

Plot measured and predicted fitness for a small design dataset. Compare score error with errors in candidate ranking. Then create a policy for choosing which additional candidates deserve expensive evaluations and how the final choice will be verified.

Sources

Yaochu Jin, A Comprehensive Survey of Fitness Approximation in Evolutionary Computation: https://www.tech.dmu.ac.uk/~syang/TF-ECiDUE/Jin-SC05.pdf. Institutional record: https://openresearch.surrey.ac.uk/esploro/outputs/journalArticle/A-comprehensive-survey-of-fitness-approximation/99512902002346. The timing calculation and FlowCraft scenario are original illustrations, not reported performance results.

Previous: Fisher's method Next: Fitness function