Model Sharing Bayesian Optimization for Mechanical Wi-Fi Antenna Control in Multi-Station Environments

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

1

Model Sharing Bayesian Optimization for Mechanical Wi-Fi Antenna Control in Multi-Station Environments

arXiv:2610.41469v1 [cs.NI] 30 Oct 2026

Akihito Taya, and Yuuki Nishiyama, and Kaoru Sezaki, Member, IEEE

Abstract—Some commercial Wi-Fi access points used in homes and offices have antennas whose orientations can be adjusted manually. However, selecting suitable orientations is difficult for non-expert users, particularly in indoor multipath environments where reflections and polarization effects complicate the relationship between antenna orientation and link quality. We investigate automatic mechanical antenna orientation control for uplink Wi-Fi communication when the set of active stations changes over time. Building on our previously developed antenna-control system, we propose a model-sharing Bayesian optimization (BO) framework that learns the log-capacity of each link using a separate Gaussian process regression (GPR) model. The framework reuses these models across active-station (STA) combinations and optimizes proportional fairness, equivalently maximizing the geometric mean of the active links’ capacities. Exploiting the linearity of this objective in log-capacity, we also derive a upper confidence bound (UCB) acquisition function and its gradient. We evaluate the framework through simulations driven by measured channel capacities and experiments with a mechanical Wi-Fi antenna prototype. The results demonstrate that the proposed model-sharing BO achieves better optimization performance than independent state-specific optimization by reusing the models trained on previously observed links. Index Terms—Mechanical antenna, automatic antenna orientation tuning, channel state information, Bayesian optimization, model sharing framework

Fig. 1: Mechanical Wi-Fi antenna system with two independently controllable antennas. The device automatically adjusts the 3D orientation of each antenna to optimize channel capacity.

I. I NTRODUCTION NTERNET of Things (IoT) networks increasingly use WiFi to carry uplink data from cameras, light detection and ranging (LiDAR) sensors, and other connected devices. As these devices start and stop their data streams independently, the set of active STAs associated with a common access point (AP) changes over time. Maintaining fair link performance across such changes is challenging even when the stations remain stationary. Indoor measurements have shown that antenna tilt and rotation can affect received signal strength through changes in radiation patterns and polarization alignment [1]. Thus, antenna orientation is a physical parameter that may be optimized alongside software-configurable radio settings. However, antennas on commercial APs typically have fixed or manually adjustable orientations, and selecting favorable angles is difficult for non-expert users in multipath-rich indoor environments. Software-controlled mechanical actuation offers a way to tune

I

A. Taya is with Institute of Industrial Science, The University of Tokyo, Meguro-ku, Tokyo 153-8505, Japan. E-mail: taya-a@iis.u-tokyo.ac.jp Y. Nishiyama and K. Sezaki are with Center for Spatial Information Science, The University of Tokyo, Kashiwa-shi, Chiba 277-8568, Japan. Manuscript received April 19, 2021; revised August 16, 2021. This work was supported by JSPS KAKENHI Grant Number JP23K26109, JP24K20759 and the Telecommunications Advancement Foundation.

antenna orientation based on observed link quality without requiring an explicit propagation model. Recent research on movable antennas (MAs) and rotatable antennas (RAs) has explored physical antenna reconfiguration to improve wireless communication performance [2]– [6]. Many studies consider array-element reconfiguration or model-based optimization, although whole-antenna rotation and polarization-aware approaches have also been investigated. Mechanically actuated systems such as NetBeam [7] have demonstrated orientation control but typically rely on tightly managed communication architectures. Our focus instead is the orientation of broad-beam antennas in conventional indoor Wi-Fi networks with time-varying sets of active STAs. The orientation that maximizes the capacity of one STA need not maximize proportional fairness across multiple active STAs. Consequently, the optimization objective changes as stations become active or inactive, even when the propagation conditions of individual links remain approximately stationary. Moreover, mechanical actuation is slow relative to wireless frame reception, so antenna control and capacity measurements must operate asynchronously. Practical systems must also handle occasional persistent changes in propagation conditions; continuously mobile stations and rapidly timevarying channels, however, are outside the scope of this study.

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

In our previous work, we developed the mechanical WiFi antenna system shown in Fig. 1 [8], [9]. Our initial study demonstrated the feasibility of automatic antenna orientation tuning for improving link capacity [8]. We subsequently introduced an asynchronous optimization framework that maintains independent optimizers for different active-STA combinations and selects the appropriate optimizer as the active set changes [9]. The conference framework also incorporated capacitybased detection of persistent environmental changes and optimizer reset, which we retain in this work. However, maintaining a separate BO model for each activeSTA combination can require up to 2N − 1 surrogate models for N STAs. Moreover, observations collected for one combination do not update the models of other combinations, even when they share the same links. This limits knowledge transfer across active-set changes and can lead to repeated exploration of previously observed link characteristics. To address this limitation, we exploit the fact that, under approximately stationary and interference-negligible conditions, the channel capacity of a given link can be modeled as a function of antenna orientation independently of the activity of other stations. Inspired by composite BO [10], [11], we maintain a separate GPR model for each link and update only the models of currently active links. For each activeSTA combination, the optimization objective is the mean logcapacity of its links, which is equivalent to maximizing their geometric-mean capacity. Previously learned link models can therefore be reused when a different active set is encountered, without learning a new surrogate for the entire combination. Assuming independent link-wise GPR predictive distributions, we further derive a UCB acquisition function and its gradient for this proportional-fairness objective. The contributions of this paper are summarized as follows:1 • We propose a model-sharing BO framework that updates per-link GPR models using observations from active STAs and reuses them across different active sets, reducing the number of surrogate models from up to 2N − 1 to N . • We analytically derive a UCB acquisition function and its gradient for proportional fairness by exploiting the linearity of the objective in log-capacity. • We conduct real-value-based simulations to statistically compare the performance of optimization methods across multiple active-set sequences. • We evaluate the proposed framework through real-world experiments using our mechanical Wi-Fi antenna prototype under controlled changes in station activity, demonstrating the benefits of reusing link models when a new active-STA combination is encountered. II. R ELATED W ORK Beamforming is a conventional approach to controlling radio propagation by steering antenna directivity mechani1 An earlier version of this work was presented at IEEE GLOBECOM 2026 [9]. The asynchronous architecture, state-specific optimization, and environment-change detection and reset mechanism were introduced in that version; the present paper extends it with per-link model sharing and a derived acquisition function.

2

cally or electronically [12]–[14]. Pal et al. [15] proposed a conformal beam-steering antenna electronically controlled by a Raspberry Pi to switch beam directions and demonstrated improved throughput in indoor environments. Such beamsteering techniques generally aim to provide directional links. In contrast, our mechanical system controls the 3D orientations of broad-beam antennas (e.g., omnidirectional antennas) to serve multiple devices through a common AP. The orientation control also changes antenna polarization, which can affect channel capacity, as demonstrated in [8]. MA systems can be classified by their actuation scale into element-level control of individual radiating elements and array-level control of an antenna structure as a whole [6]. Recent studies on element-level MAs [2], [3] have demonstrated communication performance improvements through antenna position adjustment. These systems adjust individual antenna elements to improve channel conditions and beamforming performance. In particular, RAs [4], [5] have been studied for improving communication performance by adjusting the orientations of antenna elements. In addition to element-level control, whole-array control and the reorientation of stand-alone antennas have also been studied. Mechanical antenna tilting has long been studied for cellular base stations to improve coverage and capacity [16], [17]. For 3D orientation control, NetBeam [7] investigated distributed beamforming using multiple transmitters equipped with servo motors. Although NetBeam provides insights from indoor experiments, it relies on a centrally managed architecture using software-defined radios (SDRs) for tight synchronization and channel state information (CSI) feedback. Recent six-dimensional MA studies have jointly optimized antenna position and orientation [18]–[20]. Shao et al. [18] developed a channel model and a continuous optimization method for six-dimensional antenna adjustment. Their subsequent work [19] proposed methods for cases with known or unknown channel statistics and considered discrete antenna adjustments compatible with stepper-motor control. Both studies evaluated their methods through simulations. Pi et al. [20] extended joint position-and-orientation optimization to a centrally coordinated multi-AP architecture rather than a conventional Wi-Fi network. These studies use geometrybased channel models to predict how antenna movement and rotation affect channel conditions. Wei et al. [21] developed a polarization- and mutual-coupling-aware model for RAs and evaluated alternating optimization using a 3GPP channel model and experimental measurements. In contrast, we consider broad-beam antennas in indoor environments such as homes and offices. We use GPR to learn how channel capacity varies with antenna orientation and predict capacities at unmeasured orientations without requiring an explicit geometry-based propagation model. Our focus is a model-sharing BO framework that learns separate linkcapacity functions and optimizes proportional fairness as the active-STA set changes while individual link conditions remain approximately stationary.

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

3

III. P RELIMINARY: BAYESIAN O PTIMIZATION A. Basics BO [22], [23] is a black-box optimization method that iteratively selects the next evaluation point based on a surrogate model of the objective function. Let f : X → R be a blackbox function to be maximized, where X ⊆ Rd is the input space. BO constructs a surrogate model of f using GPR, which provides a predictive distribution for the function value at any input x ∈ X . Acquisition functions are then used to balance exploration and exploitation when selecting the next evaluation point. A commonly used acquisition function is the UCB criterion, defined as follows: UCB(x) = µ(x) + κσ(x),

(1)

where µ(x) and σ(x) are the mean and standard deviation of the surrogate model at point x, respectively, and κ is a hyperparameter that controls the exploration–exploitation tradeoff. Because σ(x) tends to decrease near observed inputs and remain higher in unexplored regions, (1) balances exploitation of promising regions with exploration of uncertain ones. Concretely, BO iteratively selects the next evaluation point x∗ as the one that maximizes the acquisition function; upon evaluating f (x∗ ), the surrogate model is updated with the new observation, and the process repeats until a stopping criterion is met. For continuous search spaces, differentiable acquisition functions can be optimized using gradient-based methods such as limited-memory Broyden-Fletcher-GoldfarbShanno with box constraints (L-BFGS-B) [24], [25]. Unlike population-based methods such as particle swarm optimization and genetic algorithms, BO is designed to select informative points when objective evaluations are expensive. In our antenna-orientation optimization system, evaluations cannot be performed in parallel, and each evaluation takes several seconds because of antenna movement and subsequent capacity measurement. Therefore, BO is well suited to this sequential, measurement-intensive optimization problem. B. Composite Bayesian Optimization Astudillo and Frazier [10] proposed a composite BO framework that optimizes a function composition f (g(x)) by modeling the unknown inner function using a multi-output GPR while exploiting the known outer function. The authors then extended this framework to handle networks of intermediate functions [26]. Multiple-input composite BO was also studied in [11], which introduced the acquisition function referred to as composite UCB (cUCB). Related extensions include optimization with multiple components and target values [27] and cost-aware partial evaluation of function networks [28]. Huang et al. [29] studied tensor-output BO, including the selection of subsets of outputs contributing to the objective. Because antenna-orientation optimization for multiple active STAs can be formulated as a composite optimization problem, we draw on composite BO methods. In our problem, the active set of STAs changes exogenously over time, which alters both the available component observations and the composition defining the optimization objective. Thus, different active sets yield different compositions of the same underlying

link functions, requiring the acquisition function and observed components to be adapted to the current set. Reusing surrogate models of these link functions across active sets can improve sample efficiency. IV. S YSTEM M ODEL We consider a set of STAs S that independently start and stop uplink data streams to a common AP, thereby changing the active-STA set over time. A typical use case is an indoor IoT edge network in which an AP receives high-bandwidth uplink sensor streams, e.g., data from cameras, LiDAR sensors, or robots, from multiple sensing devices. Although we focus on the uplink, the framework could also be extended to downlink scenarios if suitable per-link channel measurements are available. Unlike conventional APs with fixed antenna orientations, our proposed AP adjusts its antenna orientations θ to optimize channel capacity. Specifically, the device independently controls two antennas in both elevation θj and azimuth φj (for antenna j ∈ {0, 1}), yielding a four-degree-of-freedom configuration space: θ = (θ0 , φ0 , θ1 , φ1 ). Upon receiving data frames from STA i ∈ S at a given orientation θ, the AP measures the CSI Hi,k (θ) ∈ Cnr ×nt , where k, nr and nt denote the subcarrier index, the number of receive antennas, and the number of transmit antennas, respectively. Using this CSI, the AP computes the achievable channel capacity, which serves as a physical-layer performance metric, and optimizes θ to maximize proportional fairness among all active STAs. The multiple-input multiple-output (MIMO)-orthogonal frequency-division multiplexing (OFDM) channel capacity Ci (θ) between STA i ∈ S and the AP is calculated as [30]:    K X 1 † Hi,k (θ)Hi,k (θ) , log2 det I + Ci (θ) = ∆f N0 ∆f k=1 (2) where ∆f , K, and N0 denote the subcarrier bandwidth, the number of subcarriers, and the noise density per antenna, respectively. Additionally, I is the nr × nr identity matrix and [·]† denotes the Hermitian transpose. Note that Hi,k (θ) is treated as an effective channel matrix that includes the received signal power scaling. To balance overall channel capacity and promote fairness among users, we adopt proportional fairness [31] based on the time-averaged capacity C̄i (θ) over a window Tavg as our optimization objective. First, the geometric mean of the channel capacity of the active STAs is computed as: !1/|S| Y ĈS (θ) := C̄i (θ) , (3) i∈S

where |S| denotes the number of active STAs. We define the proportional fairness metric FS (θ) for the active stations S ⊆ S as the logarithm of the geometric mean: FS (θ) := log ĈS (θ)
1 X log C̄i (θ) . = |S| i∈S

(4)

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

4

Since the ideal antenna orientation inherently depends on which devices are active, the system aims to find the optimal configuration θS∗ := arg maxθ FS (θ) for each unique STA combination S. Practically, commercial hardware often reports a normalized CSI, H̃i,k (θ) = αHi,k (θ). Although the scaling factor α is typically undisclosed, it can be derived from the received signal strength indicator (RSSI). Because RSSI encapsulates the total signal and noise power across all subcarriers, it satisfies: RSSI =

K X

AP with Mechanical Antenna

∥Hi,k (θ)∥ + Knr N0 ∆f,

optimization loop Check

Capacity/RSSI

Wait Twait

Request

Report Data

V. M ECHANICAL A NTENNA C ONTROL S YSTEM FOR M ULTI -S TATION O PTIMIZATION

GPR Mi

DS

GPR Mi

(b) State-specific BO

DS

GPR Mi P

DS

GPR Mi

Suggestion

In our system, the active or inactive state of each STA is determined by the presence or absence of its uplink frames. In the experimental evaluation, a STA is considered inactive if no frames are received from it for a specified period.

DS

STA selector

(7)

GPR Mi

Request

.

Optimizer

DS

Suggestion



average capacity

Optimizers Combination selector

1 H̃i,k (θ)H̃i,k (θ)† α2 N0 ∆f

Capacity/RSSI

(a) Sequence diagram of the antenna tuning framework

Request

Ci (θ) = ∆f   log2 det I +

Request suggestion Set angle θ

(5)

Substituting this into (2) yields the practical capacity formula:

k=1

Capacity/RSSI active STA

where ∥ · ∥ denotes the Frobenius norm. Consequently, α is extracted as: PK 2 k=1 ∥H̃i,k (θ)∥ . (6) α2 = RSSI − Knr N0 ∆f

K X

Capacity/RSSI

Data

k=1

×

Motor Controller

Optimizer

Buffer Data

Data

2

Optimization Manager

Capacity Calculator

STA

(c) Model-sharing BO

Fig. 2: Sequence diagram of the antenna tuning framework and the developed optimization frameworks. While the statespecific optimization framework maintains independent optimizers for each active-STA combination, the proposed modelsharing framework shares per-link GPR models across different combinations.

B. Concurrent Optimization for Multiple Active Station Combinations

A. Asynchronous System Architecture In an unmanaged network such as Wi-Fi, an AP cannot prescribe the exact packet transmission times of individual STAs. In addition, mechanical antenna movement takes tens of milliseconds, whereas frame reception and capacity estimation occur on shorter timescales. We therefore employ the asynchronous architecture developed in our conference work [9], illustrated in Fig. 2. A Capacity Calculator continuously processes received frames and maintains a moving average of the channel capacity for each observable link in a shared buffer. Independently, the Optimization Manager requests a suggested antenna orientation from the optimizer, drives the actuator, and reads the capacities from the buffer for the objective evaluation. A stabilization interval Twait is imposed after each movement so that measurements taken at the preceding orientation are not used as observations for the new one. An evaluation is discarded if a required link has no valid postmovement measurement. This decoupling avoids blocking the capacity calculation while the antenna is moving, without assuming that packet arrivals are synchronized with antenna control.

We first summarize the state-specific optimization framework introduced in [9], which forms the basis of our extension. As proportional fairness may be non-convex with respect to the multiple antenna angles, we treat the objective FS (θ) as a black-box function. BO [22] is suitable for this setting because each physical evaluation incurs antenna movement and measurement delay. In the conventional implementation, an independent optimizer O[S] models the scalar objective for each nonempty active-STA set S; its next orientation is selected using the standard UCB acquisition function (1). As shown in Algorithm 1, only the optimizer associated with the current active set is used; all other optimizers retain their learned states until their corresponding combinations reappear. If any STA required for the pending measurement becomes inactive during actuation, the evaluation is skipped rather than assigned to the wrong active set. After an optimizer reaches its stopping criterion, the system revisits its best-known orientation θSbest and compares the best observed per-link capacities with the stored baselines CS,i .

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

5

Algorithm 1 State-specific optimization framework from [9] 1: while optimization is running do 2: Obtain the current active STA set S

if S = ∅ then Wait Twait and continue Select the state-specific optimizer O[S] θ ← O[S].suggest() ▷ best known if converged Move antennas to θ and wait Twait if a required link measurement is invalid then continue 8: Obtain C̄i (θ) for all i ∈ S 9: if O[S].is_converged () then 10: if a persistent capacity change is detected then 11: O[S].reset() 12: end if 13: else 14: O[S].report(θ, FS (θ)) 15: Update the state-specific best orientation and baseline 16: end if 17: end while 3: 4: 5: 6: 7:

Specifically, the relative deviation for i ∈ S is δi =

best C̄i (θSbest ) − CS,i best CS,i

.

(8)

An environmental change is declared if the deviation exceeds γth for at least one active link during nc consecutive evaluations. The corresponding optimizer is then reset for reoptimization. This detection mechanism is inherited from [9]. C. Model-Sharing Bayesian Optimization Motivation and relation to composite optimization: The state-specific approach may maintain up to 2|S| − 1 independent GPR models, one for each nonempty combination of active STAs. More importantly, an observation from one combination cannot directly update the surrogate model for another combination, even if they contain a common link. Composite BO addresses a related modeling problem by constructing surrogate models for constituent functions [10], [11]. In particular, cUCB [11] combines the individual upperconfidence estimates through a known, generally nonlinear composition function. Utilizing this concept, we developed an antenna control framework in which both the observed links and the optimization objective change with the active-STA set. This approach, which shares the underlying channel capacity across different combinations, allows the newly encountered active combination to reuse previously learned per-link models, enabling efficient optimization for the considered problem where the objective functions grow exponentially with the number of STAs. We further exploit the linear structure of (4) to derive the predictive uncertainty of the objective directly. Active-link observations and updating shared models: Let St ⊆ S be the set of active STAs at optimization round t, and define the per-link response li (θ) = log C̄i (θ). We assume

that, within a period for which the propagation environment is approximately stationary, this channel-capacity metric for link i depends on the antenna orientation but not on the identities of other simultaneously active STAs. This is appropriate for the CSI-derived physical-layer link-capacity metric considered here; it does not imply that medium-access throughput or interference-limited rates are independent of the active set. The active set determines which link capacities can be observed and which terms enter the proportional fairness objective. For each STA i, the proposed optimizer Õ maintains a dataset Di and a GPR surrogate model Mi for li . When a valid observation at θt becomes available, the datasets of the active links receive new samples as Di ← Di ∪ {(θt , log C̄i (θt ))} and the datasets for inactive links remain unchanged. Then, the surrogate models corresponding to the active links are trained by GPR using the updated datasets. Analytically derived acquisition function: We model each log-capacity function li (θ) by a Gaussian process with predictive mean µi (θ) and standard deviation σi (θ). We assume that the posterior predictive distributions of the individual linkwise Gaussian process models are independent. Under this assumption, their linear combination in (4) is Gaussian, and the predictive mean and standard deviation of the proportional fairness objective are 1 X µi (θ), (9) µS (θ) = |S| i∈S s 1 X 2 σS (θ) = σi (θ). (10) |S| i∈S

The resulting UCB acquisition function can be expressed as UCB(θ; S) = µS (θ) + κσS (θ).

(11)

We refer to (11) as linear UCB (linUCB). For comparison, applying the cUCB construction [11] to the same linear P objective yields |S|−1 i∈S (µi (θ) + βσi (θ)), where β is its exploration parameter. Both constructions employ the same per-link regression models. However, they differ in whether standard deviations are summed or the standard deviation of the aggregate objective is calculated. We compare these acquisition functions within the same model-sharing framework in Sec. VI. The gradient of the proposed acquisition function, which can be used in gradient-based optimization, e.g., L-BFGS-B, can be obtained analytically from those of the constituent Gaussian process predictions: 1 X ∇θ UCB(θ; S) = ∇θ µi (θ) |S| i∈S P κ i∈S σi (θ)∇θ σi (θ) qP + . (12) |S| 2 (θ) σ i∈S i This expression enables gradient-based acquisition maximization using gradients of the individual GPR predictions. State management and reset handling: Although each surrogate model corresponds to an individual STA, the optimization objective depends on the current set S. The optimizer therefore maintains the best proportional fairness score FSbest , its

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

Algorithm 2 Dynamic model-sharing BO 1: Initialize: reset optimizer Õ.reset(i) for all i ∈ S 2: while optimization is running do 3: Obtain the current active STA combination S 4: if S = ∅ then Wait Twait and continue

θ ← Õ.suggest(S) Move antennas to θ and wait Twait if any STA in S is inactive then continue for all i ∈ S do Calculate average channel capacity C̄i (θ) end for 11: if Õ.is_converged () then 12: if persistent capacity change is detected then 13: Õ.reset(i) for all i ∈ S 14: end if 15: else 16: li ← log C̄i (θ) for all i ∈ S 17: Õ.report(S, θ, {li }i∈S ) 18: end if 19: end while 5: 6: 7: 8: 9: 10:

corresponding orientation θSbest , and stopping status for each active set, separately from the shared models. The Optimizabest tion Manager separately stores the per-link capacities CS,i best measured at θS for environment change detection. While the counter ci tracks the number of observations for each link to determine whether link i has collected at least ninit initial observations, the stopping criterion including the maximum number of evaluations is evaluated for each active set S. Upon reaching the stopping criterion for a particular set, its stored best orientation is used to monitor the corresponding capacities as in Section V-B. We apply the environment-change detection and reset mechanism inherited from [9] to the shared-model framework. When a persistent change is detected, the models of the currently active links are reset, and the optimization states of all active-STA combinations depending on those models are invalidated. Algorithm 2 summarizes the extended optimizer operations for time-varying active-STA sets. The methods of Õ are detailed in Algorithm 3. Unlike conventional BO with a single objective function, the proposed optimizer Õ receives the current active set S and the observed log-capacities {li }i∈S . The method Õ.report() updates the dataset and GPR model for each active STA i individually, while evaluating the proportional fairness objective for the current active set S. It also maintains the best objective score FSbest and its corresponding antenna orientation θSbest for each active set. Thus, the proposed framework separates link-specific regression models, which are shared across different active sets, from the activeset-specific optimization objectives. To accommodate the active-set transitions, the other optimizer methods are also extended. The method Õ.suggest() constructs the acquisition function UCB(θ; S) from the re-

6

Algorithm 3 Methods of the optimizer Õ 1: function R EPORT Õ.report(S, θ, {li }i∈S ) 2: for all i ∈ S do 3: Update the dataset Di ← Di ∪ {(θ, li )} 4: Train the surrogate model Mi with Di 5: Increment the counter ci

end for if score is improved P for S then 1 FSbest ← |S| i∈S li best 9: θS ← θ best 10: CS,i ← C̄i (θ) for all i ∈ S 11: end if 12: end function 6: 7: 8:

1: function S UGGEST Õ.suggest(S) 2: if stopping criterion has been reached for S then 3: return θSbest

else if ∃i ∈ S : ci < ninit then return an initial orientation for an active link i with ci < ninit 6: else 7: Calculate UCB(θ; S) using (11) 8: return arg maxθ UCB(θ; S) 9: end if 10: end function 4: 5:

1: function R ESET Õ.reset(i)

Reset the dataset Di ← ∅, the surrogate model Mi , and the counter ci ← 0 3: for all S ⊆ S with i ∈ S do 4: Reset the stopping status for S 5: Invalidate the best orientation θSbest and reset FSbest best for all j ∈ S 6: Invalidate CS,j 7: end for 8: end function 2:

gression models corresponding to the current active set S, using (11), and selects the next antenna orientation accordingly. When the active set changes, the acquisition function is reconstructed using the relevant models without discarding previously learned information. Together, the per-link model updates and active-set-dependent acquisition function enable observation reuse across different active-STA combinations without maintaining a separate regression model for each combination. The proposed method reduces the number of GPR models from up to 2|S| − 1 to |S|, but it does not remove the need to track the active sets encountered during operation. However, the main benefit of the proposed approach is to reuse informative capacity observations across active-set transitions and reduce repeated exploration of links already observed.

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

7

2.2 m

Antenna 0 Antenna 1

0°

θ0

0.7 m

RX

180°

TX A

90° θ1

0° TX A

60° 60°

4.6 m

RX

φ0

Ladder with height of 1.7 m

layout b-II

layout b-I

1.3 m

90° φ1

TX A

180°

Aluminum plate on a cart

2.2 m TX B

Desk

Fig. 3: Angle definitions for the azimuth and elevation angles.

(a) layout a: small meeting room

TX B

TX B Desk

(b) layout b: class room

VI. E VALUATION A. Evaluation Criteria and Baselines We evaluated the proposed model-sharing framework for antenna-orientation optimization under changes in the set of active STAs. We first conducted real-value-based simulations to enable reproducible and extensive comparisons of optimization methods and parameter settings, as repeated realworld experiments are time-consuming and subject to channel variability. Multiple trials under reproducible conditions are also important for fair comparisons, as BO may converge to suboptimal solutions. Based on the simulation results, we selected the parameters of the optimization methods, then conducted the real environment experiments to validate the effectiveness of the proposed method in a real environment. We used the device shown in Fig. 1 both to collect data for the real-value-based simulations and to conduct the realworld experiments. The device is based on an ASUS NUC 13 Rugged, which can independently control the elevation and azimuth of two antennas using two servo motors for each antenna. The servo motors are controlled by an Arduino UNO R4 WiFi board. The elevation and azimuth angles range from 0 to 90 degrees and from 60 to 180 degrees, respectively, as shown in Fig. 3. PicoScenes [32] was installed on the proposed receiver and the transmitting devices to acquire RSSI and CSI measurements. We utilized two devices as TXs: one is an ASUS NUC 13 Rugged (TX A) and the other is an Intel NUC 12 Pro Kit (TX B). During the evaluation, these TXs periodically inject CSI measurement packets to the proposed device, which acts as a receiver. The center frequency and bandwidth of the Wi-Fi channel are set to 5.57 GHz and 160 MHz, respectively. We compared the following five methods: (i) fixed antenna orientation (Fixed), (ii) random search (Random), (iii) statespecific BO [9] (SS BO), and (iv) model-sharing framework with cUCB [11] (MS (cUCB)), and (v) model-sharing framework with linUCB defined in (11) (MS (linUCB)). Note that the optimizer optimizes the proportional fairness defined in (4) but this section presents the geometric mean of the channel capacity (3) as the evaluation metric because it is a monotonic transformation of the objective and is easier to interpret. When comparing the convergence speed of the methods, we used normalized optimality, which is defined as the ratio of

(c) photo of layout b

Fig. 4: Layouts of the measurement environments. The proposed device (RX) and two TXs (TX A and TX B) were mounted on a ladder at a height of 1.7 m and on desks at a height of 0.7 m, respectively. In layout b, the two TXs are placed at different locations in two settings: I and II.

the geometric mean of the channel capacity to the per-state optimum in real-value-based simulation. B. Real-Value-Based Simulation To compare the performance of the proposed methods, we first conduct a real-value-based simulation. Although one run of antenna-orientation optimization takes only a few minutes, conducting sufficiently many trials in a real environment is time-consuming and difficult to reproduce. Therefore, we first measured the channel capacities in grid points of the antenna orientations in a real environment, and then conducted a simulation using the measured channel capacities. The values of the channel capacity C̄i (θ) used in the simulation are obtained in the following manner. We evaluated the proposed device in two different rooms, layout a and layout b, as shown in Fig. 4. The proposed device (RX) was mounted on a ladder at a height of 1.7 m, while the two TXs (TX A and TX B) were placed on desks at a height of 0.7 m. In layout b, we measured two settings I and II, where the two STAs are placed at different locations. In the setting I, an aluminum plate with 300 mm × 300 mm × 5 mm is placed in front of the TX B to block the direct path. During data collection, the two TXs periodically transmitted measurement frames to the receiver; interference from other devices was assumed negligible. The channel capacity C̄i (θ) is measured for each STA i and each antenna orientation θ in the two layouts. θ is discretized in 15-degree increments

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

8

Elevation [deg]

2600 2400

30 60 90 180 150 120 90 Azimuth [deg]

2200

30 60 90 60

for both azimuth and elevation angles, each ranging from 60 to 180 and 0 to 90 degrees, respectively, resulting in a total of 3,969 configurations for two antennas. At each point, we measured at least three valid CSI and RSSI values for two TXs, and then calculated the average channel capacity C̄i (θ). When calculating channel capacities, we set temperature T = 300 K and a receiver noise figure of 7 dB, which are typical values for Wi-Fi receivers [33]. The noise density is calculated as N0 = kB T F , where kB and F denote the Boltzmann constant and the linear noise factor, respectively. After the measurement, we also constructed a virtual fourlink dataset by combining these layout b-I and II. Note that this merged dataset is not a real measurement of four links simultaneously, thus it does not reflect the interference between the two TXs and propagation changes due to the presence of multiple STAs themselves. However, this dataset is useful for evaluating the optimization methods when the number of STAs increases. In the merged four-link dataset, TX A and TX B measured in layout b-II are renamed as TX C and TX D, respectively, while those measured in layout b-I are treated as they are. In these simulations, we formulate a discrete optimization problem that maximizes the geometric-mean capacity ĈS (θ) of the active STAs over the measured orientations. Although the optimal orientation can be found by exhaustive search in this discrete optimization problem, we limit the number of searches for each method to compare the optimization efficiency. Note that in the real-value-based simulation, the maximization of the acquisition function is performed by evaluating the acquisition function at all grid points of the measured antenna orientations, rather than using gradientbased optimization methods, which is a different setting from the following real experiments. Fig. 5 shows the distributions of the measured channel capacities over all antenna-orientation configurations in the grid.

60 90 180 150 120 90 Azimuth [deg]

60

−44 −45 −46 −47 −48 −49

Elevation [deg]

Fig. 5: Distribution of measured channel capacities over all antenna-orientation configurations in the grid for TX A, TX B, and their geometric mean. The boxes represent the IQR, with the median indicated by a horizontal line. The whiskers extend to the minimum and maximum observed values, and individual points represent all measured configurations without outlier exclusion.

30

30 60 90 60

(c) TX B → Antenna 0 −36

30

−38 −40

60

−42

90 180 150 120 90 Azimuth [deg]

−44

0

−36

30

−38 −40

60

30 60 90 180 150 120 90 Azimuth [deg]

60

(g) TX D → Antenna 0

−42

90 60

(e) TX C → Antenna 0 0

90 120 150 180 Azimuth [deg]

−44 −45 −46 −47 −48 −49

(d) TX B → Antenna 1

0

60

−42 −43 −44 −45 −46 −47 −48

(b) TX A → Antenna 1

Elevation [deg]

Geo. mean

90 120 150 180 Azimuth [deg]

0

90 120 150 180 Azimuth [deg]

−44

(f) TX C → Antenna 1 −41 −42 −43 −44 −45 −46 −47

0 Elevation [deg]

TX B

Elevation [deg]

TX A

Elevation [deg]

0

1800 1600

60

0

(a) TX A → Antenna 0

2000

Elevation [deg]

Capacity [Mbit/s]

2800

−42 −43 −44 −45 −46 −47 −48

Elevation [deg]

0

3000

30 60 90 60

90 120 150 180 Azimuth [deg]

−41 −42 −43 −44 −45 −46 −47

(h) TX D → Antenna 1

Fig. 6: Contour plots of the measured RSSI values from each STA to each antenna. Dots indicate the measured grid points. Different STAs exhibit distinct spatial patterns, and the optimal antenna orientations for each STA are not necessarily the same. Although the interquartile range (IQR) is only 97.6 Mbit/s, indicating that the middle 50% of configurations exhibit a relatively narrow capacity range, the geometric mean of the channel capacities ranges from 2,171 to 2,795 Mbit/s, corresponding to a 28.7% increase from the worst to the best configuration. These results demonstrate that antenna orientation can substantially affect the channel capacity, motivating the optimization of antenna orientations to identify favorable configurations for channel quality improvements and energy efficiency. The contour plots of the RSSI from each TXs to each antenna are shown in Fig. 6. It can be observed that the RSSI values vary across different TXs even for the same orientation, indicating that the optimization of antenna orientations for multiple links is a nontrivial task. In addition, the two antennas exhibit different spatial patterns. Their separation, on the order of half a wavelength, can produce different multipath fading responses; this complexity motivates our use of black-box optimization rather than an explicit geometric propagation

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

2800 2750 2700

TX A

TX B

(a) Comparison with β for cUCB

Capacity [Mbit/s]

2900

kappa 0.5 kappa 1.96

Fixed

n 20 2900

2850 2800 2750 2700

Both

n 10

TX A

TX B

2800 2750 2700 2650

kappa 1.0 kappa 3.0

Fixed 2600

2800 2750 TX A

TX A

TX B

TX B

Both

2900

Random

MS (cUCB)

TX A

TX B

2800 2750 2700 2650

SS BO

2300 2200 TX B

SS BO

MS (linUCB)

2850

2600

Fixed

MS (linUCB)

2400

TX A

Random

MS (cUCB)

Both

(b) layout a [{A}:10, {B}:10, {A,B}:10]×5

2500

2100

Fixed

Both

(b) Comparison with ninit for linUCB (a) layout a [{A}:50, {B}:50, {A,B}:50]

2850

2700

SS BO

MS (linUCB)

2850

2600

Both

Random

MS (cUCB)

2600

Capacity [Mbit/s]

2850

n5

Capacity [Mbit/s]

2900

Capacity [Mbit/s]

beta 2.0

Capacity [Mbit/s]

beta 1.0

Capacity [Mbit/s]

Capacity [Mbit/s]

2900

beta 0.5 beta 3.0

9

SS BO

MS (linUCB)

TX A

TX B

2500 2400 2300 2200 2100

Both

Random

MS (cUCB)

Both

(c) Comparison with κ for linUCB

Fixed 2900

Random

MS (cUCB)

(d) layout b-I [{A}:10, {B}:10, {A,B}:10]×5

SS BO

2800 2700 2600 2500 2400 2300

Fixed

MS (linUCB)

2900

Capacity [Mbit/s]

model. Fig. 7 shows the comparison among hyper parameters. In this evaluation, only TX A was active for the first 50 iterations, followed by TX B alone for the next 50 iterations. After 50 evaluations with only TX B active, both TXs became active, and their geometric-mean capacity was evaluated for another 50 iterations. We denote this sequence [{A}:50, {B}:50, {A,B}:50]. The parameters not compared are fixed as ninit = 5 and κ = 1.96. Although β = 2.0 shows stable performance for TX A in Fig. 7(a), the median values are comparable. Therefore, we set β = 1.0, which is the default value in [11] for the following evaluations. For linUCB, Figs. 7(b) and 7(c) demonstrate small deviation with initial points ninit = 5 and κ = 1.96 especially for TX A; thus we used these values for the subsequent evaluations. Fig. 8 compares the optimization results for the geometric mean of the channel capacity with various methods. For the fixed-orientation baseline, we show the best values among four antenna orientations θ = (θ0 , φ0 , θ1 , φ1 ), where θ represents the elevation angle and φ represents the azimuth angle. These orientations are (0◦ , 90◦ , 0◦ , 90◦ ), (90◦ , 90◦ , 90◦ , 90◦ ), (90◦ , 180◦ , 90◦ , 180◦ ), and (0◦ , 180◦ , 0◦ , 180◦ ). We evaluated two sequences: [{A}:50, {B}:50, {A,B}:50] and [{A}:10, {B}:10, {A,B}:10]×5; in the latter, each of the three active sets is evaluated for 10 iterations, and the cycle is repeated five times. Although state-specific BO outperformed random search, its results for the both-TX-active state exhibited a larger IQR than those of the model-sharing methods in layout a and layout b-II. In contrast, the model-sharing methods more consistently identify high-capacity antenna orientations. This improvement is attributed to the reuse of link observations collected during both single-TX and dual-TX active states. In

(c) layout b-I [{A}:50, {B}:50, {A,B}:50]

Capacity [Mbit/s]

Fig. 7: Comparison of parameters for cUCB and linUCB. The boxes represent the IQR, with the median indicated by a line inside each box. The whiskers follow Tukey’s 1.5×IQR rule, and outliers are plotted as individual points.

TX A

TX B

(e) layout b-II [{A}:50, {B}:50, {A,B}:50]

Both

Random

SS BO

MS (cUCB)

MS (linUCB)

TX A

TX B

2800 2700 2600 2500 2400 2300

Both

(f) layout b-II [{A}:10, {B}:10, {A,B}:10]×5

Fig. 8: Comparison of optimization methods. The boxes, lines inside the boxes, and individual points represent the IQR, median values, and outliers. The proposed model-sharing framework achieved larger capacities in all layouts. Two acquisition functions cUCB and linUCB show comparable performances.

layout b-I, where the direct path to TX B is blocked by the aluminum plate, the optimized capacities for TX B exhibit a wider distribution, with a median value substantially lower than the best observed capacity. This suggests that the antenna orientation optimization is more difficult for blocked path than line of sight (LOS) path. Fig. 9(a) and Figs. 9(b)-9(d) show the distributions of the measured channel capacities over all antenna-orientation configurations in the grid and the optimization results for three sequences2 using the merged four-link dataset. As in the twolink simulations, the model-sharing framework demonstrates greater robustness than state-specific BO, while the two acquisition functions achieve comparable performance. Even in the four-link optimization, where the number of active STAs 2 sequence α: [{A,B,C,D}:20, {A,B}:15, {B,C}:15, {C,D}:15, {A,D}:15, {A,C}:15, {B,D}:15, {B,C,D}:10, {A,C,D}:10, {A,B,D}:10, {A,B,C}:10]; sequence β: [{A}:20, {B}:20, {C}:20, {D}:20, {A,B}:20, {C,D}:20, {A,B,C,D}:30]; sequence γ: [{A}:6, {B}:6, {C}:6, {D}:6, {A,B,C,D}:6]×5

10

Normalized best-so-far

Capacity [Mbit/s]

3000 2800 2600 2400 2200 2000 1800 1600

TX A TX B TX C TX D Geo. mean

Fixed

SS BO

MS (cUCB)

0.94 Random SS BO MS (cLCB) MS (linUCB)

0.93 0.92 0.91 0.9 140

142

144

MS (linUCB)

2500 2400 2300 2200

C

D

B+ A+

A+

B+

+D

D

+D

A+ C

B+ C

B+

C

D

A+

C

A+

+D

C B+

A+

B+

C

A+

+D

B

2100

Random

MS (cUCB)

SS BO

Fixed

MS (linUCB)

2900

Capacity [Mbit/s]

2800 2700 2600 2500 2400 2300 2200

Random

MS (cUCB)

SS BO

2800 2700 2600 2500 2400 2300 2200

Random SS BO MS (cLCB) MS (linUCB)

0.96 0.95

0.94 120 125 130 135 140 145

Iteration

(b) sequence β Active TX: A, B, C, D

1 0.99 0.98 0.97 Random SS BO MS (cLCB) MS (linUCB)

0.96 0.95 0.94 20

40

60

80 100 120 140

A+

B+

C

+D

D

C

B

A

A+

B+

C

C

+D

+D

B A+

D

C

B

Active TX

(c) sequence β

0.97

Fig. 10: Optimization trajectories for the merged four-link dataset. The lines represent the mean best-so-far values over 100 trials. The metric is the normalized geometric mean of the channel capacity, where the normalization is performed by the per-state optimum. The regions corresponding to the specific active TX sets are extracted from the entire optimization trajectory.

MS (linUCB)

2100 A

2100

148

0.98

(c) sequence γ Active TX: A, B, C, D

(b) sequence α Fixed

146

0.99

Iteration

Active TX

Capacity [Mbit/s]

0.95

1

(a) sequence α Active TX: A, B, C

2600

2900

0.96

Normalized best-so-far

Capacity [Mbit/s]

Random

0.97

Iteration

(a) Measured capacity distributions

2700

0.98

Normalized best-so-far

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

Active TX

(d) sequence γ

Fig. 9: Comparisons of optimization methods for the merged four-link dataset. The boxes and lines inside the boxes represent the IQR and median values. In (b), (c), and (d), the outliers are plotted as individual points. In sequence α, the capacities of all four TXs are observed during the first 20 iterations, allowing the model-sharing frameworks to reuse the learned models when subsets of these links become active. In sequence β and sequence γ, the model-sharing frameworks exhibit a similar trend to that in the two-TX evaluations, achieving more consistent results than state-specific BO.

sets become large, the proposed model-sharing framework can efficiently conduct the optimization in unseen active STA combinations by reusing the surrogate models learned. The trajectories of the optimization results for the merged four-link dataset are shown in Fig. 10. The lines represent the average values of the best-so-far values among 100 trials, where the metric is the geometric mean of the channel capacity normalized by the per-state optimum. In sequence α, the four TXs are evaluated at the first 20 iterations, then various combinations of active TX sets are evaluated. As shown in Fig. 10(a), at the last 10 iterations, where TX A, B, and C are active, the proposed model-sharing frameworks leverage

the trained surrogate models from the previous active TX sets, and thus achieve higher capacities than random search and state-specific BO that start the optimization process from scratch. Here, the acquisition functions cUCB and linUCB show comparable performance. Fig. 10(b) shows the last 30 iterations of the optimization results for sequence β, where all four TXs are active. The figure demonstrates the effectiveness of the proposed model-sharing framework as we discussed earlier. Fig. 10(c) shows the results for sequence γ, where the active TX sets are cyclically changed. Even in frequently changing scenarios, the proposed model-sharing frameworks can efficiently optimize the antenna orientations and achieve higher capacities than the conventional methods. C. Experimental Evaluation in Real Environment This subsection presents the evaluation results of the developed antenna orientation optimization device in a real environment. The experiments were conducted in room b using the same devices as in the real-value-based simulations. Specifically, TX A was placed at the position labeled TX A in setting I, whereas TX B was placed at the position labeled TX A in setting II, as illustrated in Fig. 4(b). CSI measurement frames were injected from the two TXs to the proposed device at a period of 100 msec. The proposed device calculated the average channel capacity C̄i (θ) over Tavg = 1.0 s after

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

11

Trial 0

Trial 1

20

40

Trial 2

Trial 3

80

100

60

Capacity [Mbit/s]

Capacity [Mbit/s]

D. Discussion and Limitations 3400 3200 3000 2800 2600 2400 2200 2000 0

3400 3200 3000 2800 2600 2400 2200 2000 0

Trial 0

Trial 1

20

40

Iteration

Trial 1

20

40

Trial 2

Trial 3

80

100

60

80

100

60

(b) State-specific BO

Iteration

(c) Model-sharing BO with cUCB

Capacity [Mbit/s]

Capacity [Mbit/s]

Trial 0

Trial 3

Iteration

(a) Random search 3400 3200 3000 2800 2600 2400 2200 2000 0

Trial 2

3400 3200 3000 2800 2600 2400 2200 2000 0

Trial 0

Trial 1

20

40

Trial 2

Trial 3

80

100

60

Iteration

(d) Model-sharing BO with linUCB

Fig. 11: Optimization trajectories of the channel capacity in the real environment experiments. The plotted points and solid lines represent the measured geometric mean of the channel capacity and the best-so-far values, respectively. Four trials are conducted for each method, and the results of each trial are shown in different colors. The vertical dashed lines indicate the transitions of the active TXs. First, second, and third sections of the sequence correspond to the single-TX A, single-TX B, and both-TXs optimization, respectively.

waiting 2.0 s for the antenna actuators to settle. Active TXs were changed according to the sequence [{A}:30, {B}:30, {A,B}:50]. Active and inactive states were emulated using a software packet filter that blocked measurement frames from inactive STAs at the receiver. The proposed device considered a TX inactive if no measurement frames were received from it for 1.0 sec. We compared the four methods: random search, statespecific BO, model-sharing BO with cUCB, and modelsharing BO with linUCB. Each method was evaluated for 4 trials. The trajectories of the optimization results are shown in Fig. 11. The plotted points represent the measured geometric mean of the channel capacity at each iteration, while the solid lines indicate the best-so-far values. First and second sections of the sequence correspond to the single-TX A and single-TX B optimization, respectively, while the third section corresponds to the both-TXs optimization. Random search can occasionally find favorable orientations, so some trials achieve relatively high best-so-far capacity despite many unproductive evaluations. State-specific BO maintains independent models for the three active sets and must learn a new model when both TXs become active; many measured capacities in this segment lie between 2,500 and 2,600 Mbit/s. In contrast, model-sharing BO reuses models trained during the single-TX segments and more frequently evaluates orientations with capacities between 2,600 and 2,800 Mbit/s.

The proposed system optimizes channel capacity estimated from CSI, rather than higher-layer throughput. Actual throughput is also affected by factors such as the modulation and coding scheme (MCS), contention among multiple STAs, and transport- and application-layer protocols. We assume that the STAs share the same basic service set (BSS) and transmit separately, and therefore do not explicitly model inter-STA interference. Thus, the optimized metric of the proposed system represents the potential quality of the communication link, which is expected to improve the throughput. The framework could be applied to higher-layer throughput if the per-link objective can be measured and is sufficiently independent of the active set. When link throughput depends on interactions among active STAs, additional models accounting for these dependencies would be required; this remains future work. VII. C ONCLUSION We developed a mechanical antenna-orientation optimization device to improve channel capacity for multiple STAs whose activity changes over time. The proposed system uses a model-sharing BO framework to reuse per-link surrogate models across different active-STA combinations. Measurementbased simulations and real-world experiments demonstrated the effectiveness of model sharing relative to state-specific BO. Future work includes extending the framework to mobile STAs and moving obstacles, which introduce time-varying propagation conditions beyond the active-set changes considered here. Such extensions may require dynamic BO methods capable of tracking time-varying Gaussian process models. Another direction is to extend the mechanical antenna device to support both communication and Wi-Fi sensing. Applying the device to sensing will require selecting an optimization objective that appropriately reflects sensing performance. ACKNOWLEDGMENT The authors used OpenAI’s ChatGPT to assist with English translation, proofreading, and stylistic refinement of the manuscript. The authors reviewed and revised the resulting text and take full responsibility for the manuscript’s content. R EFERENCES [1] G. Hilton, E. Mellios, D. Kong, D. Halls, and A. Nix, “Evaluating the effect of antenna tilt and rotation on antenna performance in an indoor environment,” in 2011 Loughborough Antennas & Propagation Conference (LAPC), 2011, pp. 1–5. [2] L. Zhu, W. Ma, and R. Zhang, “Modeling and performance analysis for movable antenna enabled wireless communications,” IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6234–6250, Jun. 2024. [3] L. Zhu, W. Ma, B. Ning, and R. Zhang, “Movable-antenna enhanced multiuser communication via antenna position optimization,” IEEE Trans. Wireless Commun., vol. 23, no. 7, pp. 7214–7229, 2024. [4] B. Zheng, T. Ma, C. You, J. Tang, R. Schober, and R. Zhang, “Rotatable antenna enabled wireless communication and sensing: Opportunities and challenges,” IEEE Wireless Commun., pp. 1–8, 2025. [5] B. Zheng, Q. Wu, T. Ma, and R. Zhang, “Rotatable antenna-enabled wireless communication: Modeling and optimization,” IEEE Trans. Commun., vol. 74, pp. 6825–6842, 2026. [6] B. Ning, S. Yang, Y. Wu, P. Wang, W. Mei, C. Yuen, and E. Björnson, “Movable antenna-enhanced wireless communications: General architectures and implementation methods,” IEEE Wireless Communications, vol. 32, no. 5, pp. 108–116, 2025.

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021

[7] C. Bocanegra, K. Alemdar, S. Garcia, C. Singhal, and K. R. Chowdhury, “NetBeam: Networked and distributed 3-D beamforming for multiuser heterogeneous traffic,” in 2019 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), Newark, NJ, USA, Nov. 2019, pp. 1–10. [8] A. Taya, Y. Nishiyama, and K. Sezaki, “A mechanical Wi-Fi antenna device for automatic orientation tuning with Bayesian optimization,” in 2026 IEEE 23rd Consumer Communications & Networking Conference (CCNC), Las Vegas, NV, USA, Jan. 2026, pp. 1–2. [9] ——, “A mechanical antenna for improving capacity fairness in dynamic multi-station scenarios,” arXiv preprint arXiv:2609.16877, 2026. [10] R. Astudillo and P. Frazier, “Bayesian optimization of composite functions,” in Proc. of the 36th International Conference on Machine Learning (ICML), vol. 97. PMLR, Jun. 2019, pp. 354–363. [11] K. Jain, K. Prabuchandran, and T. Bodas, “Bayesian optimization for function compositions with applications to dynamic pricing,” in International Conference on Learning and Intelligent Optimization. Springer, 2023, pp. 62–77. [12] I. Uchendu and J. R. Kelly, “Survey of beam steering techniques available for millimeter wave applications,” Progress in electromagnetics research B, vol. 68, pp. 35–54, 2016. [13] M. I. Abbasi, M. H. Dahri, M. H. Jamaluddin, N. Seman, M. R. Kamarudin, and N. H. Sulaiman, “Millimeter wave beam steering reflectarray antenna based on mechanical rotation of array,” IEEE Access, vol. 7, pp. 145 685–145 691, 2019. [14] S. Kutty and D. Sen, “Beamforming for millimeter wave communications: An inclusive survey,” IEEE Commun. Surveys Tuts., vol. 18, no. 2, pp. 949–973, 2016. [15] A. Pal, A. Mehta, H. Goonesinghe, D. Mirshekar-Syahkal, and H. Nakano, “Conformal beam-steering antenna controlled by a Raspberry Pi for sustained high-throughput applications,” IEEE Trans. Antennas Propag., vol. 66, no. 2, pp. 918–926, 2018. [16] J. Niemela and J. Lempiainen, “Impact of mechanical antenna downtilt on performance of WCDMA cellular network,” in 2004 IEEE 59th Vehicular Technology Conference, vol. 4, Milan, Italy, May 2004, pp. 2091–2095. [17] F. Athley and M. N. Johansson, “Impact of electrical and mechanical antenna tilt on LTE downlink system performance,” in 2010 IEEE 71st Vehicular Technology Conference, Taipei, Taiwan, May 2010, pp. 1–5. [18] X. Shao, Q. Jiang, and R. Zhang, “6D movable antenna based on user distribution: Modeling and optimization,” IEEE Trans. Wireless Commun., vol. 24, no. 1, pp. 355–370, 2025. [19] X. Shao, R. Zhang, Q. Jiang, and R. Schober, “6D movable antenna enhanced wireless network via discrete position and rotation optimization,” IEEE J. Sel. Areas Commun., vol. 43, no. 3, pp. 674–687, 2025. [20] X. Pi, L. Zhu, H. Mao, Z. Xiao, X.-G. Xia, and R. Zhang, “6D movable antenna enhanced multi-access point coordination via position and orientation optimization,” IEEE Trans. Wireless Commun., vol. 25, pp. 915–930, 2026. [21] J. Wei, S. Song, Z. Kang, T. Wang, H. Gao, S. Shen, and X. Chen, “Polarization- and mutual-coupling-aware rotatable antennas enabled wireless communications: Modeling, orientation optimization, and prototyping,” IEEE Trans. Wireless Commun., vol. 25, pp. 18 119–18 135, 2026. [22] R. Garnett, Bayesian optimization. Cambridge University Press, 2023. [23] B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. De Freitas, “Taking the human out of the loop: A review of Bayesian optimization,” Proceedings of the IEEE, vol. 104, no. 1, pp. 148–175, 2015. [24] C. Zhu, R. H. Byrd, P. Lu, and J. Nocedal, “Algorithm 778: L-BFGSB: Fortran subroutines for large-scale bound-constrained optimization,” ACM Transactions on mathematical software (TOMS), vol. 23, no. 4, pp. 550–560, 1997. [25] A. Georgiou, D. Jungen, L. Kaven, V. Hunstig, C. Frangakis, I. Kevrekidis, and A. Mitsos, “Deterministic global optimization of the acquisition function in Bayesian optimization: To do or not to do?” arXiv preprint arXiv:2503.03625, 2025. [26] R. Astudillo and P. I. Frazier, “Bayesian optimization of function networks,” in Advances in Neural Information Processing Systems, vol. 34. Curran Associates, Inc., 2021, pp. 14 463–14 475. [27] J. Jeong and H. Shin, “Bayesian optimization for a multiple-component system with target values,” Computers & Industrial Engineering, vol. 157, p. 107310, 2021. [28] P. Buathong, J. Wan, R. Astudillo, S. Daulton, M. Balandat, and P. I. Frazier, “Bayesian optimization of function networks with partial evaluations,” in Proc. of the 41st International Conference on Machine Learning (ICML), vol. 235, Vienna, Austria, Jul. 2024, pp. 4752–4784.

12

[29] J. Huang, H. Xu, J. Guo, m. jiang, and C. Zhang, “Combinatorial bandit Bayesian optimization for tensor outputs,” in International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil, Apr. 2026, pp. 44 648–44 676. [30] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 21, no. 5, pp. 684–702, 2003. [31] H. Shi, R. V. Prasad, E. Onur, and I. Niemegeers, “Fairness in wireless networks: Issues, measures and challenges,” IEEE Commun. Surveys Tuts., vol. 16, no. 1, pp. 5–24, 2014. [32] Z. Jiang, T. H. Luan, X. Ren, D. Lv, H. Hao, J. Wang, K. Zhao, W. Xi, Y. Xu, and R. Li, “Eliminating the barriers: Demystifying Wi-Fi baseband design and introducing the PicoScenes Wi-Fi sensing platform,” IEEE Internet Things J., vol. 9, no. 6, pp. 4476–4496, 2022. [33] S. Merlin et al., “TGax simulation scenarios,” Nov. 2015, doc.: IEEE 802.11-14/0980r16.

Akihito Taya received the B.E. degree in electrical and electronic engineering from Kyoto University, Kyoto, Japan in 2011, and the master and Ph.D. degree in Informatics from Kyoto University in PLACE 2013 and 2019, respectively. From 2013 to 2017, PHOTO he joined Hitachi, Ltd., where he participated in HERE the development of computer clusters. From 2019 to 2022, he was an assistant professor at Aoyama Gakuin University. He has been an assistant professor at Institute of Industrial Science, the University of Tokyo since 2022. He received the IEEE VTS Japan Young Researcher’s Encouragement Award in 2012, the IEICE Young Researcher’s Award in 2018, and the IEEE CCNC Best Demo Award - Runner Up in 2026. His current research interests include IoT, distributed machine learning, and Wi-Fi sensing. He is a member of the IEEE and ACM.

Yuuki Nishiyama is an Associate Professor at Center for Spatial Information Science (CSIS), the University of Tokyo. He obtained an M.S. (2014) in Media and Governance from Keio University and PLACE a Ph.D. in Media and Governance (2017) from the PHOTO same institution. He had worked at Keio University HERE in Japan and the University of Oulu in Finland as a post-doctoral researcher, respectively. He began his career at the Institute of Industrial Science, University of Tokyo, as a Research Associate in 2019 and has worked as an Assistant Professor (Lecturer) at CSIS since 2022. Since 2025, he has held his current position. His current research interests include ubiquitous computing systems, mobile and wearable sensing platforms, and human ability augmentation. He is a member of ACM, IEEE, and the Information Processing Society of Japan (IPSJ).

Kaoru Sezaki is a professor of Center for spatial Science at the University of Tokyo and co-appointed as professor of Institute of Industrial Science, the University of Tokyo. He served as director of Center PLACE for Spatial Information Science from 1998 to 2024. PHOTO He is a steering member of e-Health Technical ComHERE mittee COMSOC. He has been general chair and TPC Chair of many IEEE international conferences. He also served as Treasurer of IEEE Tokyo Section as well as that of Japan Council from 2003 to 2004. He received B.Eng., M.Eng., and Ph.D. degrees from the University of Tokyo, Tokyo, Japan, in 1984, 1986, and 1989, respectively, all in Electrical Engineering. Since 1989, he has been with the University of Tokyo. He was a Visiting Researcher at University of California at San Diego in 1996. His research interests include e-Health, sensor networks, IoT, and urban computing.

Related documents

• Strategic Demand-Planning in Wireless Networks: Can Generative-AI Save Spectrum and Energy?#426358
• Survey on Beyond Diagonal RIS-Enabled 6G Wireless Networks: Fundamentals, Recent Advances, and Challenges#1203756
• Joint Mode Selection and User Association for ISAC-Enabled Cell-Free MIMO Networks#256611
• Wireless communications through reconfigurable intelligent surfaces#955544
• Impact of Antenna Position Errors on TDMA and NOMA in Pinching-Antenna Systems#1105646