Dynamic Batch Bayesian Optimization

Updated 23 March 2026

Dynamic batch Bayesian optimization is a technique that adaptively selects both the number and locations of query points to efficiently optimize expensive black-box functions.
It leverages methods such as DPP-based diversification, kernel quadrature, and multi-point acquisition designs to balance exploration and exploitation under varying resource constraints.
Empirical evaluations show that dynamic batching can reduce wall-clock time by up to 78% while maintaining optimization performance comparable to sequential approaches.

Dynamic batch Bayesian optimization (BO) encompasses algorithmic methodologies for optimizing expensive black-box functions by adaptively selecting batches of input queries—determining both their locations and batch sizes—at each BO round. These approaches leverage parallel evaluations to accelerate end-to-end optimization while seeking to preserve the statistical efficiency of sequential BO. Dynamic batch BO includes mechanisms to adaptively adjust the batch size or batch composition as a function of current model uncertainty, computational resource constraints, architectural combinatorics, or data-collection logistics, rather than relying on fixed batch sizes or purely single-point updates. Methodological innovations span acquisition function design, combinatorial optimization, kernel-quadrature relaxations, probabilistic inference in measure space, and DPP-based diversification. This article surveys the key principles, algorithmic frameworks, optimization procedures, and empirical findings underlying dynamic batch Bayesian optimization, drawing on foundational and recent literature.

1. Foundational Problem Formulation

Dynamic batch Bayesian optimization formalizes the objective as follows: given a black-box function $f\colon \mathcal{X} \to \mathbb{R}$ , the goal is to select a sequence of batches $\{\mathcal{B}_t\}_{t=1}^T$ (where $\mathcal{B}_t \subset \mathcal{X}$ and $|\mathcal{B}_t| = p_t$ may vary dynamically) so as to minimize the simple regret—i.e., the difference between the maximum observed $f(x)$ and the true maximum. At each round, a surrogate probabilistic model, typically a Gaussian process (GP), is fit to all collected observations, and a batch selection strategy is applied based on an acquisition function or surrogate acquisition law. Critical elements are the dynamic determination of batch size $p_t$ and the adaptive selection of batch locations, which may involve constraints, stochastic sampling, or combinatorial structures (Azimi et al., 2011, Yang et al., 2019).

Dynamic batching is motivated by practical constraints: (i) wall-clock reduction via concurrent evaluation, (ii) cost or resource limits, (iii) model accuracy considerations, and (iv) combinatorial or stochastic item generation as in high-throughput experimental domains (Yang et al., 2019). Failure to adapt batch size can degrade optimization performance due to insufficient feedback between batch selections, particularly when function responses are highly informative or non-stationary.

2. Key Methodological Strategies for Dynamic Batch Selection

Several algorithmic paradigms have emerged for dynamic batch BO:

2.1 Sequential Independence and Simulated Feedback

Azimi et al. and Gong et al. propose methods in which candidate batch points are incrementally added if the estimated influence of earlier batch elements on the GP posterior at subsequent candidate points is sufficiently small (Azimi et al., 2011, Azimi et al., 2012). This relies on quantifying the posterior mean and variance change at candidate locations, given simulated (e.g. mean or maximum possible) outcomes for earlier batch points:

Variance reduction can be computed exactly, independent of predicted outcomes.
The expected mean-shift at a candidate due to unknown previous batch outcomes is bounded, and a threshold $\epsilon$ is set to determine whether an additional point can be safely added in the batch without violating the exploit/explore trade-off.

Such schemes enable dynamic batch sizes $p_t$ , automatically adapting between fully sequential and highly parallel modes according to the model’s uncertainty and dependence structure.

2.2 Multi-point Acquisition Function Optimization

Batch acquisition functions such as multi-point expected improvement ( $\mathrm{EI}_q$ ) or multi-point UCB ( $\mathrm{UCB}_q$ ) seek to generalize single-point acquisition to the batch case (Marmin et al., 2015). Analytic gradients of $\mathrm{EI}_q$ and efficient maximization procedures enable gradient-based optimization of batch locations. Hybrids combining UCB-based candidate generation with EI-based local refinement have been shown to yield favorable empirical performance and can be coupled with dynamic batch size logic; for instance, by monitoring marginal improvement or information gain per candidate and stopping when a threshold is reached.

2.3 Diversified Sampling via DPPs

Determinantal Point Processes (DPPs) are employed to enforce diversity in batch selection, crucial when batch points selected independently tend to cluster, wasting resources (Kathuria et al., 2016, Nava et al., 2021). DPP-BBO produces a batch by sampling subsets from a DPP kernel derived from the GP posterior (e.g., based on mutual information), reweighted by an acquisition-related distribution (e.g., Thompson sampling max-distribution). The DPP mechanism naturally supports variable batch size: $k$ -DPPs can be sampled for any $k$ , and entropy-based stopping conditions allow batch sizes to adapt to the model’s state, e.g., terminating when information gain or DPP entropy falls below a prescribed threshold.

2.4 Kernel Quadrature and Probabilistic Measure-based Optimization

SOBER conceptualizes dynamic batch selection as a kernel-quadrature problem, seeking sparse quadrature rules that match moments of a belief measure over maximizer location while maximizing acquisition function values (Adachi et al., 2023). Batch size is dynamically chosen to ensure the approximation error in integration (worst-case error in an RKHS) falls below a threshold or a fixed reduction in GP posterior variance is achieved. This convex optimization-based selection is computationally efficient even for large batches and mixed input spaces.

Recent work on Wasserstein gradient flows (WGFs) lifts batch BO to the space of probability measures over the design/product space, optimizing an entropic-regularized expected utility functional (e.g., EIG) (Sharrock, 12 Mar 2026). This distributional formulation admits product-measure (i.i.d. or mean-field) relaxations to accommodate large batch sizes and dynamic batch size selection via particle-based methods. The measure’s support or product cardinality can itself be adaptively varied, e.g., by tracking the marginal information gain per batch member or by defining stopping criteria.

2.5 Submodular and Combinatorial Constraint Methods

In settings where explicit selection of batch locations is infeasible—such as combinatorial library generation in protein engineering—dynamic batch BO can operate by selecting constraints $C$ that specify a set-valued library, from which batches are sampled stochastically (Yang et al., 2019). The batch objective (expected number of improved items in the batch) is approximated and optimally decomposed into the difference of submodular functions (DS), which enables efficient, greedy local search via modular–modular (ModMod) procedures. Constraints are reoptimized at each batch based on current observations, with problem structure dictating feasible batch sizes.

3. Theoretical Performance Guarantees

Dynamic batch BO methods benefit from several nontrivial theoretical results:

Independence-based batching: Under bounded mean-shift, the blurring between sequential and parallel choices is controlled, and optimization performance matches or closely approximates that of the fully sequential policy (Azimi et al., 2011, Azimi et al., 2012).
Regret bounds: Batch versions of UCB, DPP-BBO, TS, and DPP-TS have simple and cumulative regret bounds scaling as $O(\sqrt{T \beta_T \gamma_{TB} / B})$ , where $\beta_T$ is an exploration parameter and $\gamma_{TB}$ the maximum information gain for $TB$ observations (Kathuria et al., 2016, Nava et al., 2021). DPP-based schemes can obtain strictly stronger regret bounds (by a constant offset) compared to independently sampled batches due to the repulsive, diversity-promoting property of DPPs.
Density-ratio and classification-based BO: BORE and BORE++ provide O( $\sqrt{T} \gamma_T$ ) regret bounds for distributional and simple regret under mild conditions, using density-ratio estimation (classification) with UCB-style uncertainty control (Oliveira et al., 2022).
Kernel-quadrature error and convergence: In quadrature-based approaches, moment-matching guarantees rapid decay of RKHS norm error with batch size, and the GP posterior variance reduction is tightly controlled (Adachi et al., 2023).
Particle-based flows in measure space: Wasserstein gradient flows provide contraction/convergence rates under entropic regularization, and product-measure relaxations ensure scalability with batch size; error is decomposed into finite-sample, discretization, and gradient-oracle components (Sharrock, 12 Mar 2026).

4. Algorithmic Implementation and Practical Considerations

A broad collection of algorithmic patterns exists for practically realizing dynamic batch BO:

Batch construction heuristics: Incrementally build batches by simulating outcomes or using “hallucinated feedback”; terminate based on mean-shift, marginal gain, or entropy threshold (Azimi et al., 2012, Marmin et al., 2015, Azimi et al., 2011).
Batch acquisition maximization: Employ scalable analytic gradients of batch acquisition criteria (e.g., $\mathrm{EI}_q$ ), hybridize with UCB for candidate generation, and combine with local penalization (Marmin et al., 2015, Chen et al., 2022).
Diversity and exploration control: DPP-SAMPLE and DPP-MAX batch selection, entropy-based batch size selection, quadrature error thresholds, and repulsion-based drift dynamics in particle measures support robust exploration (Kathuria et al., 2016, Nava et al., 2021, Adachi et al., 2023, Sharrock, 12 Mar 2026).
Computational complexity: Modern approaches rely on fast eigen-decomposition (DPP), convex programs (kernel quadrature), scalable gradient flows (particle–measure), and avoid combinatorial explosion even at batch sizes up to 100–200 (Adachi et al., 2023, Teufel et al., 2024).
Integration of domain constraints: Selection over constraint sets in combinatorial domains, difference-of-submodular (DS) local search, and modular bounds for greedy optimization (Yang et al., 2019).
Dynamic batch-size adaptation: Information-theoretic or utility-based heuristics to determine batch cardinality per round, supporting both problem-driven and resource-driven batch scheduling; e.g., stopping batch growth when marginal benefit falls below threshold (Azimi et al., 2011, Adachi et al., 2023, Teufel et al., 2024).

5. Empirical Evaluation and Application Domains

Empirical results across dynamic batch BO literature consistently demonstrate:

Performance parity or improvement with sequential BO in terms of simple regret, with wall-clock speedups proportional to concurrent resource utilization (up to 78% speedup reported, with negligible performance loss) (Azimi et al., 2012, Azimi et al., 2011).
Robust convergence and superior final quality in large-scale or combinatorial optimization, especially when using DS-based, DPP, or kernel-quadrature frameworks, outperforming heuristic and naive batch methods (Yang et al., 2019, Kathuria et al., 2016, Adachi et al., 2023).
Scalability to high-throughput or massive-batch regimes, e.g., batches of size 100–200 in Bayesian optimization and quadrature (Adachi et al., 2023, Teufel et al., 2024).
Domain-specific successes in protein engineering, simulation-based inference, black-box hyperparameter optimization, and sensor scheduling. DS-based combinatorial constraint optimization enables efficient search over exponential libraries and rapid identification of rare, high-fitness sequences in protein design (Yang et al., 2019); kernel-quadrature dynamics accelerate simulation-based inference and discrete space search (Adachi et al., 2023).
Effectiveness of dynamic strategies: DPP-BBO and quadrature error-based batch sizing provide adaptive batch cardinality, improving information coverage and reducing uncertainty efficiently (Nava et al., 2021, Adachi et al., 2023).
Optimization under constraints: Dynamic batch BO supports inclusion of combinatorial, discrete, or stochastic constraints in batch selection, expanding applicability beyond continuous domains (Yang et al., 2019, Adachi et al., 2023).

6. Comparative Table of Key Dynamic Batch BO Methods

Method	Batch Sizing	Acquisition/Selection Rule	Notable Features	Representative Papers
Independence-based (DBBO/Hybrid)	Dynamic (variance/mean-shift threshold)	Sequential EI/UCB + batch build-up	Provable mean-shift control, up to 78% speedup	(Azimi et al., 2011, Azimi et al., 2012)
Multipoint EI/UCB	Fixed or dynamic (via marginal EI)	Joint $\mathrm{EI}_q$ (analytic gradient)	Gradient-based maximization, hybrid heuristics	(Marmin et al., 2015)
DPP-BBO, DPP-TS	Arbitrary per-round (entropy-based)	DPP/MCMC sampling with mutual information kernel	Diversity, theoretical regret improvement	(Kathuria et al., 2016, Nava et al., 2021)
SOBER (Kernel Quadrature)	Iterative error-based	Quadrature rule (LP/QP moment-matching)	Scalable to large batch, mixed domains	(Adachi et al., 2023)
DS–Combinatorial	Problem-dependent (constraints)	DS-based greedy local search	Combinatorial constraint libraries, submodularity	(Yang et al., 2019)
Wasserstein Gradient Flow	Dynamic (measure support size)	Distributional EIG, mean-field or i.i.d.	Particle flows, scalable, contraction guarantees	(Sharrock, 12 Mar 2026)
Energy-Entropy (BEEBO)	Arbitrary, analytic control	Free-energy acquisition (explore/exploit)	Analytic gradients, heteroskedasticity support	(Teufel et al., 2024)
BORE/BORE++ (Density-ratio)	Adaptive (SVGD/uncertainty)	Classification-based surrogate, UCB	Theoretical regret, scalable, classifier-based	(Oliveira et al., 2022)

7. Perspectives and Future Directions

Dynamic batch Bayesian optimization continues to evolve, leveraging advances in surrogate modeling, combinatorial optimization, information-theoretic acquisition, and scalable inference. Open directions include:

Incorporation of non-Gaussian/non-Euclidean input spaces and integration with simulator-in-the-loop frameworks.
Tightening regret, convergence, and worst-case error bounds, especially in high-uncertainty and resource-constrained domains.
Expanding expressiveness via measure-theoretic and distributional optimization, as in Wasserstein flows and product-measure relaxations (Sharrock, 12 Mar 2026).
Automatic and theoretically sound scheduling of batch size as a function of exploration, marginal utility, or resource cost, uniting information-theoretic, statistical, and computational perspectives.
Further empirical benchmarking in diverse real-world settings, including drug/discovery, sensor networks, and automated machine learning pipelines.

Dynamic batch Bayesian optimization, through principled adaptation of batch size and batch selection, provides a robust and flexible framework for parallel, efficient optimization of expensive black-box objectives, with a rich theoretical foundation and wide empirical applicability.