Does entropy search guarantee the global optimum after a few trials?

No. It is an information-directed strategy with model and approximation assumptions, not a finite-budget guarantee for arbitrary functions.

Can it handle noisy observations?

Yes, the probabilistic model can include observation noise, but the noise specification affects what is considered informative.

Is it the same as expected improvement?

No. Expected improvement values anticipated objective improvement, whereas entropy search values information about the optimum's location.

AI Glossary / Probabilistic machine learning / Advanced

Entropy search

Definition

Entropy search is a Bayesian optimization strategy that selects an expensive function evaluation for its expected information about the location of the optimum. It aims to reduce uncertainty about which input is best, rather than choosing solely by predicted function value or uncertainty at the evaluation point.

What it means

Imagine tuning a machine through costly trials. One setting currently looks best, but another experiment might reveal whether a whole competing region is better. An informative trial need not itself produce an excellent result; it may eliminate uncertainty that prevents a good final decision.

Entropy search maintains a probabilistic belief about the objective and, from it, a belief about the optimizer's location. Entropy summarizes how spread out that belief is. The method asks how much a proposed observation is expected to concentrate it. It is not ordinary search-engine retrieval, nor the entropy-minimization penalty used to sharpen classification predictions.

How it works

The original method models the unknown objective with a Gaussian process and approximates the distribution of its global minimizer over a bounded search domain. For a candidate evaluation, it considers possible observation outcomes under the current predictive distribution.

Each hypothetical outcome would update the model and the belief about the minimizer. The acquisition value reflects the expected reduction in uncertainty. The next experiment is selected by optimizing this acquisition function, then the actual observation updates the model.

Computing the optimizer distribution and averaging over future outcomes is difficult, so practical entropy search uses approximations. Related methods, including predictive entropy search, exploit alternative formulations of the information objective. Their names should not imply identical numerical algorithms or interchangeable implementation details.

Three practical examples

1. An engineering team tunes a few simulator parameters when every simulation takes hours. It considers which run would most clarify the best configuration within its allowed range.

2. A model-development team selects a costly training configuration based on information about the best hyperparameters, while accounting for variability between training runs.

3. A laboratory compares controlled process settings under a fixed experimental budget. The information criterion guides candidate selection, while separate constraints determine which trials are permissible.

Think of it this way

Think of looking for the lowest pass through a mountain range. Measuring one ridge may rule out several possible routes even if that ridge itself is high. The analogy stops at the need for a probabilistic model: entropy search estimates how informative a measurement will be, rather than knowing in advance what it will reveal. An incorrect model can make an apparently useful measurement disappointing.

Worked example

Use a deliberately simplified discrete example with two possible optimizer locations, A and B. Current probabilities are 0.5 and 0.5, so entropy is one bit.

Candidate experiment X has two equally likely outcomes. Under the supplied model, one outcome changes the optimizer probabilities to (0.9, 0.1), and the other to (0.1, 0.9). Each posterior has entropy −0.9 log2(0.9) − 0.1 log2(0.1), approximately 0.4690 bits. Expected information gain is therefore 1 − 0.4690 = 0.5310 bits.

Candidate Y instead produces posterior probabilities (0.6, 0.4) or (0.4, 0.6), again with equal outcome probabilities. Its expected posterior entropy is approximately 0.9710 bits, so its expected gain is 0.0290 bits. With equal evaluation costs, this criterion prefers X.

The average posterior probabilities for either candidate return to (0.5, 0.5), consistent with the stated prior. These supplied discrete beliefs illustrate information-gain arithmetic, not a Gaussian-process implementation. They do not specify which experiment has the better immediate objective value or prove that either will find the true optimum.

Illustrative case study

Fictional case: Resin Harbour has a limited budget for testing production settings in a simulator. Two regions look promising, but the existing measurements cannot distinguish them. The analyst uses an entropy-search acquisition to identify a trial expected to clarify their relative merits.

The selected trial returns a mediocre objective value yet reduces uncertainty about the competing region. The team evaluates the final recommended setting separately and checks sensitivity to the surrogate's noise assumptions. It does not count informative but poor trials as production successes. This is an illustrative workflow, not a reported industrial result.

Common misconceptions and mistakes

Entropy search is not equivalent to choosing the point with the largest predictive variance. Uncertainty in a region that cannot plausibly contain or inform the optimum may be less useful than a smaller uncertainty affecting the final choice.

Nor does it simply choose the currently most probable optimizer location. Sampling elsewhere can be more informative. Expected information gain is an average over possible observations; an individual trial can produce less gain than anticipated or even make current beliefs more diffuse.

A concentrated posterior is also not proof that the optimum has been found. Model assumptions, numerical approximations and excluded regions can create unjustified confidence.

Business use

The business value comes from using scarce evaluations efficiently when the final configuration matters more than the reward of every exploratory trial. The method is most attractive when experiments are expensive relative to the computation required to choose them.

Define search bounds, measurement noise, trial cost and final decision criteria. Compare against simpler acquisition strategies under the same evaluation budget. Where costs vary, account for them explicitly rather than assuming information gain alone captures business value. Keep feasibility and operational constraints visible, and use a final confirmation run to evaluate the recommended setting under realistic conditions.

Check your understanding

In the worked example, does preferring experiment X establish that A is the optimum?

Answer: No. X is preferred because its possible outcomes are expected to distinguish A from B more clearly. Either outcome remains possible before the experiment.

Related terms

Bayesian optimization
Information gain
Gaussian process
Expected improvement
Global optimization
acquisition function
predictive entropy search
experimental design

Learn next

Compare entropy search and expected improvement on a small candidate set. Track separately the best observed value, the final recommendation and uncertainty about the optimizer, since they measure different aspects of progress.

Expected improvement

Sources

Primary sources: Hennig and Schuler, Entropy Search for Information-Efficient Global Optimization: https://jmlr.org/papers/volume13/hennig12a/hennig12a.pdf ; Hernandez-Lobato and colleagues, Predictive Entropy Search, primary abstract: https://arxiv.org/abs/1406.2541 . Independent review checked original setting and Section 2.5 information objective. The two-experiment example is an original discrete illustration, not a reproduction of either algorithm.

Previous: Entropy rate Next: Entry-sequenced organization