Bayesian Optimization over Windows (BOW)

• Bayesian Optimization over Windows (BOW) is a framework that restricts Bayesian optimization to admissible regions (windows) instead of a single point optimum.
• It applies to diverse problems by modeling tolerance regions in property space, sampling regions in design, or reachable control sets in motion planning.
• BOW leverages Gaussian process surrogates and specialized acquisition functions to balance exploration, feasibility, and diversity within dynamic constrained settings.

Searching arXiv for the cited BOW-related papers to ground the article in current literature. I’m going to look up the relevant arXiv records for BOW and closely related “window/range-aware” Bayesian optimization papers. Bayesian Optimization over Windows (BOW) denotes a class of Bayesian optimization formulations in which the optimization target is defined by a window, region, or short-horizon admissible set rather than by unconstrained extremization of a scalar objective. In recent arXiv literature, the term is used explicitly in motion planning for constrained optimization over a dynamic planning window of reachable controls (Raxit et al., 18 Aug 2025) and for short-horizon steering inside a bidirectional kinodynamic planner (Raxit et al., 25 Jun 2026). Closely related work uses different terminology—most notably range-aware Bayesian optimization, target range, tolerance-aware acquisition, specification-driven design, and Bayesian sampling windows—for problems in which the goal is to recover valid designs inside a prescribed property window or to identify near-optimal sampling regions rather than a single exact design point (Jiang et al., 10 Jun 2026, Buchhorn et al., 2022).

1. Conceptual scope and terminology

BOW is not a single standardized formalism. The literature currently uses the phrase in at least two explicit senses and one closely related precursor. In the motion-planning papers, a window is a reachable local control set or receding-horizon interval over which constrained BO is performed (Raxit et al., 18 Aug 2025, Raxit et al., 25 Jun 2026). In the range-aware inverse-design paper, the corresponding object is a target range in output space, defined not as a single optimum but as an acceptable region around a target property vector (Jiang et al., 10 Jun 2026). In the Bayesian-design paper, the object is a sampling window or region that preserves high design efficiency under operational constraints (Buchhorn et al., 2022).

This terminological dispersion is technically important. It means that “window” may refer to a set in property space, a parameterized region in design space, or a short-horizon reachable control set. A common misconception is that BOW necessarily means optimization over an axis-aligned interval or box. The range-aware inverse-design formulation instead uses a Euclidean tolerance ball in $K$-dimensional property space, reducing to an interval only in the scalar case $K=1$ (Jiang et al., 10 Jun 2026). Another misconception is that every window-based method is a standard sequential BO loop with EI or UCB. The Bayesian sampling-windows paper uses a GP emulator over window parameters but does not define an acquisition function or adaptive sequential BO loop (Buchhorn et al., 2022).

Setting	Meaning of “window”	Optimization output
Range-aware inverse design	Euclidean tolerance region in output space	Valid designs satisfying a target range
Bayesian sampling windows	Parameterized sampling region	High-efficiency regions
Motion planning	Reachable control window over a short horizon	Feasible local control or trajectory segment

A plausible implication is that BOW is best understood as an organizing idea: Bayesian optimization restricted to an admissible region that is itself central to the problem formulation.

2. Range-aware Bayesian optimization for target property windows

The clearest non-robotic formulation of BOW appears in “Range-Aware Bayesian Optimization for Discovering Diverse Designs within Target Property Windows” (Jiang et al., 10 Jun 2026). The problem is inverse design over $\mathcal{X} \subset \mathbb{R}^M$ with black-box property map

$f : \mathcal{X} \to \mathbb{R}^K.$

The surrogate is a collection of $K$ independent Gaussian processes, one per output dimension, with zero prior mean and Matérn $5/2$ kernels. For output $k$, the posterior predictive mean and variance are

$$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$

$$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$

The defining shift is from extremum-seeking BO to specification-driven validity. A design is valid if its predicted properties lie within a tolerance radius $\varepsilon$ of a target vector $K=1$0: $K=1$1 When target properties differ in scale or acceptable tolerances, each output dimension is normalized before applying this Euclidean criterion. For multiple specifications, the paper defines targets

$K=1$2

with valid sets

$K=1$3

The central acquisition is Tolerance Ball (TB): $K=1$4 Using the isotropic predictive-variance approximation

$K=1$5

the paper derives the closed form

$K=1$6

where

$K=1$7

This is the mathematical core of the framework: the acquisition equals the posterior probability of satisfying the target range. In the scalar case, the formulation reduces to the probability that the scalar posterior lies inside a target interval.

The second acquisition, Heaviside (HV), strongly prioritizes points whose posterior mean already lies inside the tolerance region: $K=1$8

$K=1$9

TB is the cleanest BOW-style acquisition in this setting because it directly scores range satisfaction rather than improvement relative to a best-so-far scalar objective.

The optimization loop is conventional at the outer level—initialize with $\mathcal{X} \subset \mathbb{R}^M$0 Latin hypercube samples, fit the GP, maximize the acquisition, evaluate, update, and repeat—but unconventional in what is being optimized. For multiple specifications, the method uses a shared surrogate and shared observation set. At each outer iteration, the GP is fit once, and each target independently proposes one candidate under the same fixed posterior; duplicate proposals are resolved by a fixed ordering. This implements parallel pursuit of multiple specifications without a joint combinatorial objective.

A further distinction is that diversity is central to evaluation but not explicit in the acquisition. TB and HV are pointwise feasibility acquisitions and “do not explicitly penalize similarity to previously discovered valid candidates.” Diversity arises only implicitly from posterior uncertainty, feasible-region geometry, and the prohibition on duplicate evaluations. In continuous spaces, the paper measures diversity via $\mathcal{X} \subset \mathbb{R}^M$1-uniqueness and the normalized area under the uniqueness curve,

$\mathcal{X} \subset \mathbb{R}^M$2

while in discrete spaces it uses

$\mathcal{X} \subset \mathbb{R}^M$3

3. Bayesian sampling windows in fully Bayesian design

“Bayesian Design with Sampling Windows for Complex Spatial Processes” addresses a distinct but closely related problem: exact optimal designs are often operationally brittle because sampling at specific coordinates may be infeasible in field settings (Buchhorn et al., 2022). The paper therefore reformulates design output from exact points to windows or regions that preserve near-optimal design efficiency.

The starting point is fully Bayesian optimal design with utility $\mathcal{X} \subset \mathbb{R}^M$4, prior $\mathcal{X} \subset \mathbb{R}^M$5, and expected utility

$\mathcal{X} \subset \mathbb{R}^M$6

The optimal design is

$\mathcal{X} \subset \mathbb{R}^M$7

Because this integral is intractable, the paper uses Monte Carlo

$\mathcal{X} \subset \mathbb{R}^M$8

and often replaces the mean with the median utility

$\mathcal{X} \subset \mathbb{R}^M$9

for robustness under skewed or heavy-tailed utility distributions.

The main utility is Kullback-Leibler divergence from posterior to prior,

$f : \mathcal{X} \to \mathbb{R}^K.$0

that is, Lindley’s expected information gain criterion. Posterior inference is approximated with a Laplace approximation, because each utility evaluation requires Bayesian inference and is therefore expensive.

The paper first solves exact-point design using stochastic coordinate exchange and a nonparametric acceptance criterion based on the one-sided Wilcoxon rank-sum test. The window formulation is then introduced over a continuous domain: $f : \mathcal{X} \to \mathbb{R}^K.$1 Here $f : \mathcal{X} \to \mathbb{R}^K.$2 is the number of window parameters. In the river-network case, $f : \mathcal{X} \to \mathbb{R}^K.$3 and each coordinate is a candidate location along one of two predefined neighbourhoods. In the coral-reef case, $f : \mathcal{X} \to \mathbb{R}^K.$4 and each coordinate is a transect radius $f : \mathcal{X} \to \mathbb{R}^K.$5, with image locations perturbed by

$f : \mathcal{X} \to \mathbb{R}^K.$6

The objective over window parameters is then emulated with a GP. Training data are sampled window configurations

$f : \mathcal{X} \to \mathbb{R}^K.$7

with computed median utilities $f : \mathcal{X} \to \mathbb{R}^K.$8. The emulator assumes

$f : \mathcal{X} \to \mathbb{R}^K.$9

using the additive exponential kernel

$K$0

The nugget $K$1 is included specifically so that Monte Carlo approximations of the median utility surface are smoothed rather than interpolated. Hyperparameters are selected by minimizing the cross-validation criterion

$K$2

Prediction uses the GP posterior mean

$K$3

which is normalized into a design-efficiency measure

$K$4

where

$K$5

The output is the high-efficiency region

$K$6

Technically, this is not standard sequential BO. The method uses an offline sample/evaluate/emulate/predict workflow and thresholds the GP predictive mean to obtain efficiency contours. That distinction matters: the paper is best viewed as GP-emulated optimization over window parameters rather than as acquisition-driven adaptive BO. The window concept nevertheless serves the same broader function as in other BOW formulations: it replaces a brittle point optimum with an admissible region whose performance remains close to optimal.

4. Motion-planning BOW: dynamic control windows and constrained BO

The term Bayesian Optimization over Windows is used explicitly in two motion-planning papers, where a window is a local reachable control set or short receding-horizon interval (Raxit et al., 18 Aug 2025, Raxit et al., 25 Jun 2026). Both papers address online kinodynamic planning in cluttered environments, where random control sampling is sample-inefficient and geometric heuristics are unreliable under dynamics.

In “BOW: Bayesian Optimization over Windows for Motion Planning in Complex Environments,” the robot evolves according to

$K$7

and the full planning problem is written as constrained optimization over controls

$K$8

subject to dynamics and safety constraints

$K$9

The central simplification is to optimize inside a dynamic planning window

$5/2$0

based on current velocity and acceleration limits, and to hold the control constant over a short horizon: $5/2$1 This yields a short-horizon proxy cost

$5/2$2

abbreviated as $5/2$3. Obstacle avoidance is encoded by

$5/2$4

The method then performs constrained BO over $5/2$5. Using sampled controls $5/2$6, it builds datasets for the objective and each constraint, models them with GPs, and selects

$5/2$7

with

$5/2$8

$5/2$9

This is the motion-planning meaning of BOW in its most direct form: BO over a reachable dynamic window with explicit probabilistic feasibility.

“BOWConnect: Parallel Bayesian Optimization over Windows with Learned Local Cost Maps for Sample-Efficient Kinodynamic Motion Planning” embeds a related local BOW primitive inside a bidirectional parallel tree-based planner (Raxit et al., 25 Jun 2026). Here the state is

$k$0

with dynamics

$k$1

and rollout map

$k$2

For a candidate control $k$3, the local reward is

$k$4

and the feasibility label is

$k$5

Two GPs are maintained, one for reward and one for the constraint, producing $k$6. Feasibility is converted to

$k$7

and the acquisition is

$k$8

The control selected by this constrained local BO yields a short kinodynamically feasible motion segment inserted into a MotionTree. The full planner launches $k$9 forward workers and $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$0 backward workers, each with its own MotionTree and BOW instance. Opposite-direction trees are connected by spatial hashing,

$$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$1

followed by a kinematic pre-check and a boundary value problem bridge. In this architecture, BOW is not a global path optimizer but a learned local steering function.

5. Surrogates, acquisitions, and what is actually optimized

Across these papers, the most stable technical commonality is not a particular kernel or acquisition but a change in the semantic role of the acquisition itself. In standard BO, EI, PI, or UCB are usually tied to improving a best-so-far scalar objective. In BOW-style methods, the acquisition is defined relative to a windowed acceptance concept.

In range-aware inverse design, the acquisition is exactly the posterior probability of satisfying a Euclidean tolerance region. Under the isotropic approximation, that probability is available in closed form through the noncentral $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$2 CDF (Jiang et al., 10 Jun 2026). The baseline methods in that paper—EI, LCB, BAX, and random sampling—are all adapted to the target-discrepancy setting, but they remain discrepancy-minimizing or uncertainty-reducing methods rather than direct range-satisfaction methods. The conceptual difference is explicit: BAX explores an estimated feasible set, whereas TB maximizes the posterior probability of feasibility.

In Bayesian sampling windows, the surrogate has a different role. The GP does not drive a sequential acquisition loop; rather, it emulates a noisy utility surface over window parameters, and the predictive mean is converted into efficiency contours (Buchhorn et al., 2022). The method therefore uses a GP in the style of BO without actually instantiating modern adaptive BO. This suggests that “optimization over windows” can include both sequential and offline-emulation paradigms.

In the motion-planning literature, the objective is localized both spatially and temporally. The BOW Planner optimizes CEI over a dynamic window $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$3, with the objective and constraints evaluated by short-horizon rollout under the true dynamics (Raxit et al., 18 Aug 2025). BOWConnect does something similar but separates reward and feasibility into two GPs and uses a feasibility-weighted lower-confidence-style acquisition over the local control window (Raxit et al., 25 Jun 2026). In both cases, the window is not a state-space box and not a global path parameterization; it is the set of controls relevant to a local steering decision over a short horizon.

A technically useful way to distinguish the variants is by the object being modeled:

• Output-space validity modeling: the range-aware paper models $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$4 and scores the probability that it falls inside a tolerance ball.
• Window-parameter utility emulation: the sampling-windows paper models $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$5 over parameterized regions and thresholds normalized predictive mean.
• Local control-window optimization: the motion-planning papers model short-horizon cost/reward and feasibility as functions of control and optimize an acquisition within a reachable set.

This suggests that BOW is less a single algorithm than a structural motif: BO or GP emulation constrained by an application-specific window that encodes admissibility, flexibility, or local reachability.

6. Empirical behavior, limitations, and open directions

The empirical evidence is strongest in the range-aware inverse-design and motion-planning settings. On benchmarks including Branin, Hartmann-3, Ackley-5, Layeb-6, several pool-based materials and molecular datasets, and two application studies, the range-aware paper reports that TB achieves the best average rank across tasks and tolerance ratios, the highest fraction of best-performing tasks, and generally stronger balanced recovery across targets (Jiang et al., 10 Jun 2026). In parallel target search, it also attains low cross-target hit ratios, which the paper interprets as target fidelity. In the polymer molecular weight distribution design case study over a 51,200-point grid, TB reaches mean normalized discrete diversity $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$6, ahead of HV at $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$7, LCB at $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$8, and BAX at $$\mu_k(\mathbf{x}_*) = \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{y}_k,$$9, while EI and random are near zero. In the sequence-defined conjugated oligomer study over 1,980 unique oligomers, TB again has the highest mean normalized diversity score, $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$0, although random sampling is competitive for some targets because valid molecules are more abundant.

The Bayesian sampling-windows paper illustrates a different empirical value proposition: not better pointwise optimization, but robustness to implementation constraints (Buchhorn et al., 2022). In the river-network case, the design-efficiency contours reveal that one neighbourhood is much more sensitive to exact placement than the other, and the paper gives the example that choosing $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$1 in $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$2 can still permit a location $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$3 in $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$4 while retaining $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$5 optimal utility. In the coral-reef case, the optimal radii are

$$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$6

with reported efficiency $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$7. The paper interprets this as evidence that some transects admit substantial positional flexibility while others effectively require exact placement.

The motion-planning papers report both runtime and success-rate advantages. The BOW Planner uses only 5 samples per planning horizon in the main UGV benchmarks and identifies the optimum with 15 samples in the sample-efficiency figure (Raxit et al., 18 Aug 2025). Across six simulated $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$8 UGV environments, it consistently has the lowest total planning times and lowest per-step times while solving all six environments. In Environment 6, which contains non-convex obstacles, BOW reports path length $$s_k^2(\mathbf{x}_*) = \kappa_k(\mathbf{x}_*, \mathbf{x}_*) - \mathbf{k}_{k,*}^{\top}\left(\mathbf{K}_k + \sigma_{n,k}^2 \mathbf{I}\right)^{-1}\mathbf{k}_{k,*}.$$9 m, total time $\varepsilon$0 ms, and time per step $\varepsilon$1 ms, while HRVO and CBF produce no solution. The same paper reports five 3D UAV environments with a 12-D state and 4-D control problem, again emphasizing sample efficiency and low planning time.

BOWConnect extends this local BO steering into a bidirectional parallel kinodynamic planner and reports 100% success across ten benchmark environments, including narrow passages and non-convex spaces (Raxit et al., 25 Jun 2026). Direct comparisons to the prior unidirectional BOW planner are particularly revealing: in Narrow Passage-1 under the Unicycle model, BOW requires $\varepsilon$2 s versus $\varepsilon$3 s for BOWConnect, both at 100% success; in Narrow Passage-2, $\varepsilon$4 s versus $\varepsilon$5 s; and in Intel, BOW reaches 80% success with $\varepsilon$6 s while BOWConnect reaches 100% success with $\varepsilon$7 s. Real-world deployment on a ground vehicle and a quadrotor reports real-time planning with no collisions.

The limitations are equally domain-specific. The range-aware framework relies on factorized GPs and the isotropic predictive-variance approximation

$\varepsilon$8

which ignores output correlations and variance anisotropy; diversity is measured but not explicitly optimized (Jiang et al., 10 Jun 2026). The Bayesian sampling-windows method does not use predictive uncertainty in acquisition, depends on surrogate quality for its efficiency contours, and remains computationally heavy because each utility evaluation requires Monte Carlo and posterior approximation (Buchhorn et al., 2022). The BOW Planner is fundamentally a local/receding-horizon method, assumes conditional independence of constraints in its feasibility probability, and may get trapped in local minima such as narrow passages or bugtrap-type situations (Raxit et al., 18 Aug 2025). BOWConnect mitigates some of these issues with bidirectionality and BVP bridging, but its speedup is still limited by connection checking and sequential BVP solving, and the paper identifies adaptive worker allocation, uncertainty-aware planning beyond the current setup, and dynamic environments with moving obstacles as future work (Raxit et al., 25 Jun 2026).

Taken together, these results indicate that BOW is most coherent when understood as a specification- or admissibility-centered reformulation of Bayesian optimization. In one branch, the window is a target region in output space and the acquisition is the posterior probability of validity. In another, the window is a design region whose utility is emulated and thresholded into efficiency contours. In motion planning, the window is a short-horizon reachable control set over which constrained BO acts as a steering mechanism. The shared idea is not a single canonical implementation, but the replacement of brittle point optimality by Bayesian reasoning over acceptable regions, ranges, or local admissible windows.