Bayesian optimization is a strategy for finding the inputs that maximize or minimize an objective function when that function is expensive to evaluate, lacks a known analytical form, and may return noisy outputs. Rather than relying on gradients or exhaustive search, it builds a probabilistic model of the function from the observations it has gathered so far and uses that model to decide where to sample next. The approach is designed to be sample-efficient, meaning it tries to extract as much information as possible from each costly evaluation. This makes it well suited to problems such as tuning machine learning hyperparameters, calibrating simulators, optimizing experimental designs, and configuring control systems.
The core idea
At its heart, Bayesian optimization rests on two interacting components: a surrogate model that approximates the true objective and an acquisition function that proposes where to evaluate next. The surrogate captures both a best guess of the objective's value across the input space and an estimate of how uncertain that guess is. The acquisition function then combines these predictions with an explicit policy for trading off exploration of uncertain regions against exploitation of regions that already look promising. By alternating between fitting the surrogate to new data and optimizing the acquisition function to choose the next query point, the method gradually concentrates its evaluations on the most informative locations.
Why a probabilistic surrogate matters
The use of a probabilistic surrogate is what distinguishes Bayesian optimization from simpler black-box search methods. A purely deterministic approximation tells you only what the function might look like, while a probabilistic one also tells you how much that estimate can be trusted at each point. This uncertainty is essential because the method must reason about points it has never seen, deciding whether sampling there could plausibly reveal a better optimum. Without a calibrated notion of uncertainty, the optimizer would have no principled way to balance the risk of wasting evaluations against the potential gain of discovering improvement.
Gaussian processes as the standard surrogate
The most common surrogate is a Gaussian process, which defines a distribution over functions rather than a single function. Given a kernel that encodes assumptions about smoothness and length scales, a Gaussian process produces a posterior mean and variance at every input, conditioned on the observed evaluations. The mean serves as the prediction and the variance as the uncertainty, both of which feed directly into the acquisition function. Gaussian processes are popular because they handle small data sets gracefully, allow uncertainty to grow naturally away from observed points, and offer closed-form posterior updates under standard assumptions.
Alternatives to Gaussian processes
While Gaussian processes are the default choice, they scale poorly with the number of observations and can struggle in very high-dimensional spaces. Random forests and tree-structured Parzen estimators provide alternatives that handle conditional, discrete, or categorical variables more naturally, which is common in hyperparameter search. Bayesian neural networks and deep kernel learning offer richer function classes when the underlying objective has complex structure. The choice of surrogate ultimately depends on the dimensionality, the data budget, and the nature of the input variables.
Acquisition functions
The acquisition function turns the surrogate's predictions into a concrete decision about where to evaluate next. Expected improvement measures, in expectation, how much better a candidate point might be than the best value seen so far. The upper confidence bound combines the posterior mean with a scaled version of the standard deviation, explicitly weighting exploitation and exploration. Probability of improvement and entropy-based criteria, which target the points expected to reduce uncertainty about the location of the optimum, offer further alternatives with different theoretical guarantees and empirical behaviors.
Balancing exploration and exploitation
The exploration-exploitation tradeoff is the central tension in Bayesian optimization. Pure exploitation would always sample near the current best, risking getting stuck in a local optimum, while pure exploration would scatter evaluations widely and waste the budget. Acquisition functions encode this balance mathematically, often with a tunable parameter that shifts emphasis between the two modes. As the optimization progresses and uncertainty shrinks in promising regions, the natural behavior of well-designed acquisition functions is to gradually focus more on exploitation while still probing genuinely uncertain areas.
Handling noise and stochastic objectives
Many real-world objectives are noisy, returning different values when evaluated at the same input. Bayesian optimization handles this by treating each observation as a noisy realization of the underlying function, with the surrogate model incorporating a noise term in its likelihood. The best observed value is then replaced by the best predicted value under the posterior, since the raw observations can no longer be trusted as exact. Acquisition functions can also be adapted to account for noise, ensuring that the optimizer does not chase fortunate fluctuations.
Constraints and multiple objectives
Practical problems often involve constraints, such as latency budgets or memory limits, that must be respected alongside the primary objective. Constrained Bayesian optimization models each constraint with its own surrogate and modifies the acquisition function to weight candidates by the probability of feasibility. When multiple objectives must be balanced, the method can be extended to approximate a Pareto front, using acquisition criteria that measure expected improvement in dominated hypervolume or similar multi-objective notions. These extensions preserve the sample-efficient character of the underlying approach.
Parallel and batch evaluation
When evaluations can be performed in parallel, such as training several models simultaneously on a cluster, the optimizer must select a batch of points rather than a single one. Naively choosing the top candidates from the acquisition function tends to produce redundant samples clustered in the same region. Batch methods address this by encouraging diversity, often through techniques such as fantasized observations, where hypothetical outcomes at pending points are used to update the surrogate before selecting additional candidates. Local penalization and joint optimization of batch criteria are other strategies that promote informative parallel evaluation.
High-dimensional and structured inputs
Standard Bayesian optimization tends to degrade as the input dimension grows, because the surrogate's uncertainty estimates spread thin and the acquisition surface becomes hard to optimize. Several strategies mitigate this, including assuming an effective low-dimensional subspace, decomposing the function into additive components over groups of variables, or applying random embeddings that project the search into a smaller space. For structured inputs such as molecules, graphs, or programs, specialized kernels or learned latent representations allow the method to operate in a space where similarity is meaningful.
Practical considerations
Implementing Bayesian optimization well requires attention to several details beyond the choice of surrogate and acquisition function. Inputs are typically normalized to a common scale, kernel hyperparameters are fit by maximizing the marginal likelihood or marginalized over with priors, and the acquisition function itself must be optimized at each iteration, often using multistart local search. Initial design matters as well, with space-filling designs such as Latin hypercube sampling providing a reasonable starting set before the model-driven phase begins. Diagnostics that examine model fit and acquisition behavior help detect pathologies such as overconfident posteriors or premature convergence.
Where it fits among optimization methods
Compared to grid search and random search, Bayesian optimization usually finds better solutions with far fewer evaluations, especially when each evaluation is costly. Compared to gradient-based methods, it does not require derivatives and is robust to noise, multimodality, and discrete variables, though it generally cannot match gradient methods when those are available and reliable. Compared to evolutionary algorithms, it tends to be more sample-efficient but less straightforward to parallelize at very large scale. The right choice depends on the cost of evaluation, the structure of the search space, and the available computational resources.
Summary of how it functions
Bayesian optimization functions as a closed loop in which a probabilistic model of an unknown objective is repeatedly updated and queried. Each cycle proposes a new evaluation, observes the result, and refines beliefs about where the optimum lies. The method's power comes from making each query as informative as possible rather than from sheer computational brute force. This combination of probabilistic reasoning and decision-theoretic sampling makes it a foundational tool wherever evaluations are expensive and structure must be inferred from limited data.