WorksEngineering
Joint Movable Antenna Positioning and Precoding for ISAC via a Two-Timescale Model-Based GNN
Abstract
Movable antennas (MAs) allow an integrated sensing and communication (ISAC) base station (BS) to reshape multiuser channels and angular sensitivity. For monostatic multiuser communication and multi-target elevation and azimuth estimation, we minimize the expected target-averaged angular Bayesian Cramér-Rao bound (BCRB) subject to per-user block-level and long-term rate requirements. To avoid repeated online optimization, we propose a two-timescale model-based graph neural network (TT-MBGNN) over antennas, users, and targets. The slow policy selects positions from long-term statistics to reduce movement overhead, while the fast policy learns complex corrections to a regularized zero-forcing (RZF) reference using post-movement channels and target-angle priors. We jointly train both policies through a differentiable model of rates and BCRB, without optimization-generated labels. Across tested rate requirements, the network yields lower sensing bounds than variants using a fixed array, RZF, or both, with any-user fast-rate violations in 0.31-2.63% of test blocks. Frozen policies provide competitive sensing bounds for unseen user and target counts, and ten fine-tuning epochs improve communication reliability at higher user counts. Across tested network sizes, our framework achieves up to 3 dB lower BCRB than an idealized alternating-optimization benchmark that uses known target reflection coefficients. The network computes precoders approximately 50-fold faster in matched batch-one tests.
Introduction
Integrated sensing and communication (ISAC) supports communication and environmental sensing through shared radio frequency (RF) components, transmitted signals, and spectral resources [1, 2]. This integration can improve spectral and hardware efficiency, while the base station (BS) array geometry shapes both communication and sensing performance. Most existing systems employ fixed-position antennas (FPAs) whose element locations remain unchanged after deployment. Digital precoding controls the complex antenna excitations, while the spatial sampling positions remain set by the deployed geometry. Consequently, fixed geometries limit the flexibility to reshape multiuser channel correlations, mitigate interference, and adjust the effective sensing aperture.
Movable antenna (MA) technology addresses this limitation by allowing antenna elements or subarrays to change their positions within prescribed movement regions [3, 4]. Closely related position-reconfigurable architectures are also studied under the fluid antenna system (FAS) terminology [4]–[6]. The FAS concept includes liquid conductors, switchable ports, reconfigurable pixels, and mechanically repositioned antennas. This paper specifically considers MAs with continuously adjustable physical coordinates. Such reconfiguration can improve channel gains, reduce multiuser interference, and reshape the sensing response without increasing the number of RF chains. These benefits have motivated joint position and beamforming designs for communication systems [3, 7] and for ISAC systems based on sensing signal to interference and noise ratio (SINR), beampattern, or Cramér-Rao bound (CRB) objectives [8]–[11]. However, antenna positions enter the communication channels and sensing steering vectors through nonlinear phase terms and are tightly coupled with the precoders. The resulting nonconvex problems are commonly addressed using iterative combinations of alternating optimization (AO), successive convex approximation (SCA), majorization-minimization (MM), penalty or dual-decomposition methods, and gradient-based updates [7]–[9], [11, 12].
A further challenge arises from the different adaptation rates of mechanical positioning and electronic precoding. Moving the antennas in every coherence block can incur excessive delay and positioning overhead [10, 13]. Moreover, directly measuring instantaneous channels across many candidate antenna positions entails additional acquisition overhead. These limitations motivate slow-timescale positioning using user equipment (UE)-location distributions and long-term propagation statistics, followed by fast precoding using instantaneous channel state information (CSI) acquired at the selected geometry [10, 14]. In our formulation, the slow inputs additionally include target-prior statistics. The sensing objective must also account for block-dependent uncertainty in the target directions. The Bayesian Cramér-Rao bound (BCRB) incorporates prior information into estimation-error bounds [15]. We use this bound as a common sensing criterion, with the slow geometry evaluated across block states and the fast precoder adapted to the current prior.
Existing optimization-based studies address complementary aspects of joint positioning and beamforming. The hybrid architecture in [7] jointly optimizes movable-subarray positions and analog and digital beamformers. Slow positioning based on localized user regions and statistical CSI is studied in [10], while discrete and continuous two-timescale designs are developed in [13, 14]. A recent cell-free multiple-input multiple-output (MIMO) design also optimizes antenna positions and
A. Fadakar and A. F. Molisch are with the Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, California, USA; E-mail: {fadakarg,molisch}@usc.edu.
Y. Zhang is with the Electrical and Computer Engineering Program, Computer, Electrical and Mathematical Sciences and Engineering (CEMSE), King Abdullah University of Science and Technology (KAUST), Thuwal 23955-6900, Kingdom of Saudi Arabia; E-mail: [email protected].
This work is supported by the National Science Foundation (Grants 2504811, 2292535 and 2433981). (Corresponding author: Yuchen Zhang.) orientations using global statistical CSI, while updating decentralized beamformers on a faster timescale [16]. These studies establish the value of coordinating mechanical positioning and electronic precoding across different timescales.
Existing iterative designs typically require repeated optimization as operating conditions change, increasing online computational cost. Learning-based methods address this limitation by transferring much of the optimization effort to offline training [17], [18]. In particular, graph neural networks (GNNs) represent interactions among network entities through shared message functions and have been applied to beamforming, power allocation, and energy-efficiency optimization [19], [20]. Shared parameters also allow the same graph-layer structure to accommodate different numbers of users and targets. Heterogeneous graph architectures for cell-free and cooperative ISAC generally assume fixed antenna geometries [21]–[23].
For MA and FAS systems, learning-based studies have explored multi-agent and deep reinforcement learning, generative learning, and hierarchical policies for movement, port selection, and beamforming [24]–[27]. The graph-based designs in [28], [29] address positioning and beamforming without separating statistical geometry selection from block-adaptive precoding, while the two-timescale deep-unfolding method in [30] considers communication objectives. Supervised alternatives based on optimization-generated labels also incur the cost of repeatedly executing a numerical solver. A model-based graph neural network (MBGNN) can embed differentiable physical models into training [20], enabling direct objective evaluation without these labels. A gap therefore remains in jointly learning long-term positioning and geometry-dependent precoding for Bayesian multi-target angle estimation under communication requirements at both timescales.
Motivated by these observations, we propose a two-timescale model-based graph neural network (TT-MBGNN) for monostatic MA-assisted ISAC. The slow-timescale model-based graph neural network (S-MBGNN) selects a common geometry from long-term statistics, and the fast-timescale model-based graph neural network (F-MBGNN) adapts the precoder to current block information. Both policies are trained jointly without optimization-generated labels through a differentiable model that evaluates communication rates and target-averaged angular BCRBs. The user data streams also support sensing, without dedicated sensing streams [1], [11], [12]. Target ranges and reflection coefficients enter only the sensing model and are excluded from the policy inputs.
Main Contributions
The main contributions are summarized as follows.
We formulate a two-timescale stochastic design for monostatic MA-assisted ISAC. One continuous antenna geometry is chosen per slow interval from UE-location and channel statistics, target-angle prior statistics, and long-term rate requirements. Each block precoder adapts to post-movement CSI, current target priors, and fast-timescale rate requirements. The objective minimizes the expected target-averaged angular BCRB subject to per-UE block and long-term rate requirements, transmit power, antenna movement, and minimum-spacing constraints.
We develop a heterogeneous TT-MBGNN with an S-MBGNN for geometry selection and an F-MBGNN for block-adaptive precoding. Relation-specific messages combine attention and mean aggregation over MA, UE, and target nodes, and complex MA-UE channels affect the fast policy as edge features. The proposed decoder learns entrywise complex corrections to a regularized zero-forcing (RZF) reference and normalizes the resulting precoder to meet the power limit. Shared parameters make the graph layers equivariant to consistent node relabeling at inference and support different node counts without changing the layer parameterization.
We train both policies jointly without optimization-generated labels. A differentiable communication-sensing model evaluates block rates and angular BCRBs, providing block-specific precoder gradients and a sample-average sensing objective for the common geometry. One-sided augmented-Lagrangian losses and projected multiplier updates address fast and slow quality-of-service (QoS) requirements and raw antenna spacing. Structured outputs satisfy the power and movement-region limits. Although the proposed finite geometry repair does not provide a universal spacing guarantee, every repaired geometry checked in the reported simulations met the minimum antenna spacing requirement.
Simulations on unseen realizations demonstrate lower sensing bounds than learned-component ablations and fixed-array RZF across the tested QoS levels. Across the evaluated network-size configurations, including fine-tuned policies, the proposed framework achieves up to 3 dB lower angular BCRB than the idealized sample-average-approximation-based alternating optimization (SAA-AO) benchmark. Frozen policies provide competitive sensing bounds at unseen user and target counts, while ten additional epochs improve communication reliability at larger user counts. Batch-one per-block precoder computation is approximately 50 times faster than iterative refinement in the considered scenario.
Notation: Scalars, vectors, and matrices are denoted by regular lowercase, boldface lowercase, and boldface uppercase letters, respectively, while calligraphic letters denote sets or mappings as defined in context. The superscripts and represent transpose and Hermitian transpose, respectively, and . The operators , , and return the real part, imaginary part, and scalar magnitude. The symbols and denote the Euclidean and Frobenius norms, respectively. The identity matrix is denoted by , and is the -dimensional all-ones vector. We use for vertical stacking of scalar or vector arguments, for concatenation, and for the Hadamard product. Finally, denotes the positive-part operator, is the cardinality of set , and denotes Euclidean projection onto the set .
System and Signal Model
As illustrated in Figure 1, we consider a monostatic ISAC system in which a dual-functional BS, equipped with continuously adjustable MAs and dedicated RF chains, serves single-antenna UEs while supporting two-dimensional (2D) direction of arrival (DOA) estimation for point targets in elevation and azimuth1. The corresponding index sets are , , and . We adopt a two-timescale positioning and precoding design in which the antenna geometry is selected using long-term statistical information and held fixed over a positioning interval of transmission blocks, indexed by . Each block contains symbol intervals, indexed by . The digital precoder is updated in each block using the selected geometry and current block information. The statistical inputs remain valid throughout the positioning interval, whereas the instantaneous conditions are approximated as constant within each block and may vary between blocks. To describe spatial variability, we associate UE and target with spherical location regions
Figure 1. Considered monostatic MA-assisted ISAC with a continuous -plane aperture. Shaded spheres denote long-term location regions, and markers illustrate block-dependent UE and nominal target locations. Teal arrows represent downlink transmission, while solid and dashed purple arrows denote probing and echoes, respectively. The common geometry is selected from long-term statistics and held fixed for coherence blocks, whereas the precoder is updated per block using current CSI and target-angle priors.
where are the region centers and are their radii. These parameters specify the regional distributions, which remain unchanged within each positioning interval. In block , the UE and nominal target locations are sampled independently across users, targets, and blocks as
where denotes the uniform distribution with respect to volume over the specified region. The sampled locations remain unchanged within a block. The nominal target location determines the center of the block-dependent angular prior introduced in Section II-D. Residual transmitter leakage and hardware-induced self-interference are assumed sufficiently suppressed and are omitted from the signal model [10].
Remark 1 (Statistical inputs and temporal information). As detailed in Section IV, we provide the F-MBGNN with current block information and the S-MBGNN with long-term statistical summaries available before antenna positioning. These summaries may be estimated from historical observations. Neither network explicitly processes an ordered block sequence or maintains a recurrent state. Under the stationary model, we use independent snapshots to estimate the statistical summaries and expected training loss. We leave explicit trajectory modeling and history-dependent policy learning for future work.
Movable-Antenna Array Architecture
A local Cartesian coordinate system is considered at the BS, with its boresight aligned with the positive -axis and the continuous antenna-movement region located on the -plane. The position of the th MA is , with the corresponding three-dimensional coordinate . The complete array geometry is collected as . The movement-region boundaries, minimum inter-antenna separation, and other physical constraints are specified in the optimization formulation. Let
denote a far-field 2D DOA, where is measured from the -plane toward the positive -axis and is measured in the -plane from the positive -axis toward the positive -axis. The corresponding unit direction vector is defined as . Because the antennas move only on the -plane, define the projected direction vector as . Hence, . Under the far-field plane wave model, the position-dependent steering vector is
where is the carrier wavelength and the exponential is applied element-wise. Unlike a conventional fixed array, the manifold in (4) is directly controlled by . Consequently, moving the antennas modifies both the spatial correlations of the communication channels and the effective aperture and angular sensitivity available for target sensing.
Downlink Communication Signal Model
Position-Dependent Geometric Rician Channel
For the downlink to UE , located at in block , we adopt a narrowband far-field geometric Rician model for continuous MA coordinates [9, 10, 31]. Let denote the large-scale power gain, the distance from the BS reference point to the UE, and the line-of-sight (LoS) departure direction. The corresponding propagation phase is , and the steering vector is . The scattered field is approximated by effective non-line-of-sight (NLoS) clusters. For cluster , let denote the nominal departure direction and the aggregate gain of unresolved diffuse contributions. Independent gains are assumed across clusters.2 Each cluster is represented by the nominal steering vector . The channel vector is
where is the Rician factor in linear units, drawn from the prescribed long-term distribution. For , the angular-offset vector is drawn from , independently between elevation and azimuth. Here, is the offset standard deviation in radians, and each offset has support on . The th offset is added to to generate the corresponding cluster direction. The resulting azimuth is wrapped modulo onto the adopted azimuth interval, while elevation is restricted to the valid domain of the elevation convention. Thus, the Gaussian law specifies the offsets before these domain mappings. Since the steering vectors are explicitly parameterized by the antenna coordinate matrix , any modification to the MA geometry dynamically alters the phase of every geometric path. Ultimately, upon establishing , the BS acquires the aggregate block-level channel matrix
Transmitted and Received Signals
The UEs are served simultaneously using linear spatial division multiple access (SDMA). Let denote the precoder for UE and the block precoding matrix. The independent Gaussian data symbols are collected in , with independent draws across symbol intervals. The transmitted signal is
Conditioned on , the signal covariance and average transmit-power constraint in each block are
The same data-bearing signal is used for sensing, without additional dedicated sensing streams [11, 12]. The received signal at UE is
where is additive noise independent of the transmitted data.
Instantaneous and Long-Term Rate Metrics
With Gaussian signaling and the other users’ streams treated as noise, the block-level SINR and spectral efficiency of UE are
The rate is measured in bit/s/Hz. Both quantities remain constant within a block under the assumed channel and precoder model. To describe block-to-block randomness, define
Here, collects the aggregate diffuse cluster gains, and collects the angular offsets defined above. The sensing state and the resulting target-prior information are specified in Section II-D1. The location distribution is given in (3), while the gain and angular-offset laws are specified in the preceding channel subsection. The Rician factor is generated as
The limits and angular spreads are listed in Table I. Independent draws are used for the random communication inputs across users and blocks in the simulations. The generative state is used for sampling and performance evaluation. During deployment, the fast policy receives the resulting post-movement CSI and current target-prior information, without requiring the individual cluster gains or angular offsets. For an integrable function , the Monte Carlo estimator is obtained as:
| System and Channel | Training and Evaluation | ||
| Dimensions | MA/UE/target counts | Node/relation MLPs | two affine layers, hidden width 256, SiLU. |
| Frequency | GHz, mm, m (far-field boundary) | Optimizer / budget | AdamW, learning rate , 200 epochs, 100 steps/epoch |
| Bandwidth / power | 10 MHz, symbols/block, dBm | Input scaling | Gain offset , Rician scale ; rate scale bit/s/Hz; dropout |
| Noise | W; 290 K; communication/sensing noise figures 7 dB each | Clipping / decay | Slow/fast gradient-norm caps 0.5/1; weight decay 0 |
| Array / reference | , minimum spacing ; reference: URA over | QoS pairs | bit/s/Hz |
| Default UEs | Centers : , , m; radii : m | Margins / tails | bit/s/Hz; |
| Added UEs | Centers , m; radii m | PHR / dual steps | ; |
| Default targets | Centers : , m; radii : m | Multiplier values | Initially 0; |
| Added target | Center m; radius 10 m | Spacing repair | Targets , ; iterations, step , tolerance |
| Rician / diffuse | , clusters, ; spreads | Training / validation | statistics samples; training: blocks, angles/block; validation: 256 blocks, 64 angular samples/block, every epoch |
| Target prior / phase | Raised-cosine full widths rad, cross section m, | Final test | 200 slow intervals, blocks/interval, 96 angular samples/block |
| Gains / losses | All element/UE gains 0 dBi; system/noise losses 0 dB | Size adaptation | 10 epochs, 100 steps/epoch; learning rate |
| Network and Numerical Settings | SAA-AO | SAA-AO (Adam, learning rate 0.02) | |
| Graph layers | Node width ; slow/fast depths ; attention heads | Samples / updates | 64 blocks, 48 angular samples/block; 30 outer iterations, each with 5 geometry and 30 precoder updates |
| Decoder MLPs | Position : ; precoder : | Initialization | 20 starts: reference array and 19 layouts with independent perturbations of coordinate logits; RZF precoders |
| Output / loss | Gains ; clipping , power/log floors , sensing weight | Primal updates | Geometry/precoder gradient-norm cap 1/5 |
| Constraint terms | Rate margins 0; fast/slow/spacing penalties and dual steps $(0.01, 0.01, 0.02); multipliers initially 0, cap 100 | ||
| Test refinement | 100 precoder updates, gradient-norm cap 5, penalty coefficient 10; dual step 0.05, initial value 0, cap 100 |
Table I. Simulation and Implementation Parameters. MLP widths are listed from input to output.
where indexes independent numerical samples and indexes physical transmission blocks. The realized rate averaged over one positioning interval is . For a given geometry and a fixed rule for block-adaptive precoding, a stationary block-state process yields the ensemble-average rate and its numerical estimate as
The equality follows from the common marginal distribution and does not require independent physical blocks. The slow-timescale QoS criterion constrains this ensemble mean, without guaranteeing the realized average in every finite interval. Thus, specifies the geometry persistence in physical blocks, whereas specifies the numerical sample count.
Monostatic Multi-Target Sensing Model
Following representative MA-assisted ISAC models [10, 12, 13], each target is modeled as a far-field point scatterer with a LoS round-trip path. Clutter, target-induced multipath, external interference, and residual transmitter leakage are neglected as previously assumed. The targets are stationary within each coherence block, such that their angles and reflection coefficients remain constant over its snapshots but may vary across blocks. Doppler is therefore ignored, and block-dependent angular uncertainty intervals are assumed available from a preceding sensing or tracking stage.
Let and denote the direction and position-dependent steering vector of target in block , respectively. Since the same array is used for transmission and reception, we obtain the monostatic spatial response as . Let denote the unknown complex reflection coefficient, which absorbs the round-trip path loss, propagation delay, common phase, and radar cross section (RCS). It is treated as a deterministic nuisance parameter within each block. Defining and , the aggregate sensing channel is:
The known data-bearing signal in (7) is reused as the sensing probe. The echo received in symbol interval is obtained as
where is independent across blocks and symbol intervals. Define and . Stacking the sensing snapshots gives
where .
FIM and BCRB Analysis for 2D-DOA Estimation
We construct a separate angular BCRB for each block using a hybrid information formulation with random target angles and deterministic nuisance reflection coefficients [32]. The standard regularity and local-unbiasedness conditions for the deterministic-parameter directions are assumed. We define the block-level parameter vectors as and , where stacks the elevation and azimuth angles of all targets. Letting denote the sensing-channel derivative with respect to the -th entry of , the block-level data Fisher information matrix (FIM), accumulated over transmitted symbols and averaged over the unit-power data symbols, can be obtained via the Slepian–Bangs formula [6, 33] as
Consequently, depends on the fast-timescale precoder through and on the slow-timescale MA geometry through the target steering responses and their angular derivatives.
Block-Dependent Target-Angle Priors
We distinguish the block-to-block distribution of the target prior from the conditional distribution of the unknown target angles within a given block. The former generates diverse target contexts for offline training, while the latter governs the Bayesian sensing model. For the nominal target location sampled in (3), the angular-prior center is given by:
where maps a Cartesian location to its elevation or azimuth direction from the BS. We collect the sensing variables used for block generation and bound evaluation as
The reflection coefficients enter the outer block-state average but are excluded from both policy inputs. The prior information available at the beginning of block is collected as
Thus, the across-block distribution of is induced by the Cartesian target-location distributions through the nonlinear angle mapping in (19); the prior centers are not assumed to be uniformly distributed in angle. In the considered implementation, each target has configured support widths that remain constant throughout the slow interval, i.e., . Hence, only the prior centers vary across blocks. Conditional on , the local uncertainty around each prior center is modeled by a raised-cosine density. This compactly supported prior approaches zero smoothly at its boundaries and satisfies the regularity conditions of the Van Trees inequality [15, 34]. Defining , the conditional density of angular component is
Assuming conditional independence among targets and angular components, the joint within-block prior is obtained as . The resulting prior FIM is obtained as:
with diagonal entries ordered consistently with .
Fast-Timescale Per-Block BCRB
For a fixed geometry , precoder , and prior parameter set , the data FIM must be averaged over the joint target-angle prior corresponding to block . Because the resulting -dimensional expectation lacks a tractable closed form, it is approximated using joint angle samples:
where . The deterministic nuisance parameter vector remains fixed across all samples. To provide a differentiable mapping from the interval parameters to the sampled angles, each sampled component is generated via the affine reparameterization , where is drawn from the standardized density . Partitioning into angular and nuisance blocks, we obtain the equivalent Bayesian angular information matrix through the Schur complement, , assuming that is nonsingular. This construction accounts for the unknown reflection coefficients under the per-block nuisance model in Section II-C. The per-block BCRB matrix and the target-averaged sensing metric are defined as:
where denotes the principal submatrix associated with the elevation and azimuth angles of target .
Slow-Timescale Expected BCRB
Under stationary block states, the expected average sensing metric over the blocks of a slow interval equals the expected per-block BCRB. Using the sample-average operator in (13), the slow-timescale objective and its Monte Carlo approximation are
Here, specifies the duration of the common geometry in coherence blocks, whereas is the number of samples used to approximate the expectation.
Two-Timescale Optimization Problem Formulation
We jointly design the MA geometry and digital precoders according to their different adaptation timescales. The geom etry is selected from long-term statistics at the beginning of each slow interval, with mechanical repositioning completed before block transmission. At the beginning of block , a control stage provides the post-movement CSI and the target-prior information . An external tracker or predictor supplies the latter using observations available before block . Acquisition and precoder computation are assumed to finish while both inputs remain valid for the ensuing communication-sensing symbols. For , define the requirement vector . A fast policy maps the selected geometry, current channel, target-prior information, and fast requirements to a precoder:
Here, denotes a general block-adaptive mapping, whose neural-network parameterization is introduced in Section IV.
Fast-Timescale Digital-Precoder Problem
For a fixed feasible geometry , we formulate the model-based blockwise precoder-design problem as
The objective is the block-level BCRB in (25), evaluated using and the block’s fixed nuisance realization . Simulated nuisance values are used for offline training and bound evaluation and are excluded from the learned policy inputs. For compactness, denotes the rate in (10) evaluated with . The problem is nonconvex because the precoding vectors are coupled through multiuser interference and the sensing information matrix.
Slow-Timescale MA-Position Problem
The common geometry is selected before the future block states are known. Each candidate is therefore evaluated through the fast policy . For Monte Carlo sample , let denote the output of (27) using and . Both the channel and the resulting precoder depend on the candidate geometry, even when the sampled environment is unchanged. Thus, is determined by and and it is not treated as an independent optimization variable. We write and for the estimators in (26) and (14), respectively, evaluated using these geometry-dependent channels and policy-generated precoders.
Following the continuous-position MA model in [10, 35], define as the movement region and as the minimum inter-MA spacing. For a fixed fast policy , the sample-average geometry problem is
The last two constraints enforce the fast-timescale requirements on the sampled states. Satisfaction of the rate requirements on unseen states is assessed separately in the numerical evaluation. The thresholds and specify block-level and long-term service requirements, respectively, and need not be equal. Although expresses the geometry subproblem for a fixed fast policy, the proposed framework jointly trains the slow geometry policy and . Consequently, both policies account for the sensing objective and the fast- and slow-timescale communication requirements.
Proposed Two-Timescale Model-Based GNN
Architecture and Information Flow
The proposed TT-MBGNN separates slow-timescale geometry selection from block-adaptive precoding while training the two decisions jointly. As illustrated in Figure 2, the S-MBGNN selects one MA geometry per positioning interval from long-term information. Given this geometry, the F-MBGNN produces one precoder in each of the following coherence blocks using the current block information. The corresponding policy mappings are represented as:
Figure 2. Architecture of the proposed TT-MBGNN. The S-MBGNN selects one geometry per positioning interval, and the F-MBGNN produces one precoder per block. Final MA states feed the geometry decoder, while final MA and UE states feed the precoder decoder. Target information reaches both decoders through message passing. Channel and RZF coefficient features also enter the pairwise decoder. Joint offline training uses the differentiable physics model and raw-spacing loss, with a straight-through estimator (STE) for finite repair. Solid and dashed red arrows denote forward computation and training gradients. The policies have the same layer structure and separate parameters.
Here, are the raw and repaired geometries, respectively, and denotes the finite spacing-repair procedure. The vectors and contain the trainable parameters of the two policies.
Slow-Timescale Input Features
Slow-Timescale Input Features: The input to the S-MBGNN is collected as
The matrix contains the normalized reference coordinates of the MAs, where the reference geometry is . Each row of describes one UE through its long-term departure direction, large-scale channel gain, Rician factor, and slow-timescale requirement. Hence, is already included in . Finally, each row of encodes the nominal elevation and azimuth prior centers and their angular widths for one target. Please refer to Appendix A for the explicit slow-feature definitions and normalization.
Fast-Timescale Input Features
Fast-Timescale Input Features: For block , the F-MBGNN input features are collected as
The matrix contains the normalized antenna coordinates of the selected geometry . Each row of summarizes the current channel magnitudes and includes the corresponding fast-timescale rate requirement. The matrix encodes the current target-angle prior centers and widths from . The edge-feature tensor provides three real-valued features for each complex coefficient in . The same feature vector is assigned to both directed MA-UE edges associated with the corresponding link. Please refer to Appendix A for the explicit feature definitions and normalization.
During training, the differentiable communication-sensing model evaluates the achieved rates and angular BCRBs and supplies gradients to both policies. This model-based procedure requires no optimization-generated labels [36]. During deployment, fixed-depth policy evaluations and output processing replace repeated optimization of the physical variables.
Heterogeneous Graph and Relation-Aware Layers
Both policies operate on the directed heterogeneous graph , containing MA, UE, and target nodes. The five relation types are given by:
Each relation is a complete directed bipartite connection between its source and destination types. For relation and destination , the incoming neighborhood is . The two MA-UE relations also carry channel edge features in the F-MBGNN. For policy , let be the column-vector feature of node , with dimension , from (32) or (33). A policy- and type-specific encoder initializes . The state is then updated through relation-aware layers. In the equations below, we fix policy and, for the fast policy, coherence block , and write . Consequently, and are the destination and source states at the same layer; both are encoder outputs when and outputs of layer otherwise. Relation-specific message functions distinguish the physical interactions among the node types [37]. Attention permits different source nodes to contribute unequally [38], while parallel mean aggregation provides a neighborhood summary normalized by the number of source nodes. The gated residual update subsequently combines these messages without discarding the node’s previous state. For and , the relation-specific Multi-Layer Perceptron (MLP) computes
Here, denotes concatenation. The feature has for the two fast MA-UE relations and is omitted otherwise (). Let divide , set , and partition each message into head vectors . For a learned matrix , . Normalizing separately over the sources of each relation gives
The attention and mean aggregates, followed by their relation-specific fusion, are
where the aggregates and belong to , and is an MLP. Attention weights sum to one within each destination, relation, and head before training-time dropout. Let be the relations entering nodes of type . A type-specific affine map processes the node’s own state. The type-specific MLP then combines that state with the relation outputs:
An affine map forms the element-wise gate
where is the logistic function. The layer output is
where is type-specific layer normalization. Dropout is disabled during inference. Please refer to Appendix B for the affine parameterizations and training-time dropout. At inference, the graph layers are equivariant to consistent relabeling within each node type, with node and edge features permuted accordingly [36]. This claim excludes the index-dependent direction rule for nearly coincident antennas used by the spacing repair in Section IV-C1.
Structured Policy Decoders
After the relation-aware layers, the slow MA states determine the array coordinates. The fast MA and UE states jointly determine each precoder coefficient. Information from target nodes reaches these decoders through the directed graph relations.
Slow-Timescale Geometry Decoder
After the graph layers, a shared position MLP maps the final MA state to a correction relative to . Let be the coordinate bounds of . Define and , where and the operations are element-wise. Column of is
where controls the correction magnitude and is the element-wise logistic function. This mapping keeps the raw coordinates within .
As will be detailed in Section IV-D, training includes a dedicated penalty that encourages separation between MAs. Since this penalty does not guarantee feasible spacing for every decoded layout, we additionally apply a finite geometric correction in each forward pass during training and deployment. The correction acts on antenna pairs with distances below before the fast policy and communication-sensing model are evaluated. Define and initialize . At iteration , let . For , use . For nearly coincident positions, use the fixed unit vector , where and is the pair index in a fixed ordering. The proposed pair corrections and simultaneous coordinate update are given by:
where , clips coordinates to , is the update scale, and counts incident pairs whose distance shortfall exceeds . Each pair proposes equal and opposite shifts of half the shortfall, and each antenna averages the incident corrections. For an isolated violating pair with and no boundary clipping, reaches the spacing target in one update. With multiple pairs, simultaneous corrections and clipping can leave residual violations. The procedure stops when all distances are at least or after iterations. The returned geometry is . The fast policy and sensing and rate losses use the repaired geometry, so the forward performance effects of these distance-based corrections are included in joint training. During training, a STE passes gradients to the raw coordinates [39]:
where preserves its argument in the forward pass and has zero derivative. The resulting identity backward derivative is a surrogate for the derivative of the finite repair.
Remark 2 (Spacing feasibility). Although the finite repair does not provide a universal spacing guarantee, we verified from the unsmoothed coordinates that every repaired geometry checked in § 5 had a minimum spacing of at least .
Fast-Timescale Precoder Decoder
The fast decoder adds learned complex corrections to an RZF reference. For , define
where is a numerical constant. The resulting transmit power does not exceed . Writing , the reference RZF precoder is given by:
where , and controls the regularization. For each MA–UE pair , the shared decoder takes the final MA and UE states, the channel feature , and a three-dimensional feature of the corresponding RZF coefficient. The two outputs form the real and imaginary parts of . The block precoder is
where controls the correction magnitude. The entry-wise corrections can change the RZF beam directions, while final normalization enforces the power constraint. Please refer to Appendix C for the coefficient feature and pairwise correction function.
Physics-Informed Constrained Training
At each training step, the selected geometry is evaluated over independently sampled block states . For each state, the fast policy produces , and the differentiable physics model estimates the angular BCRB . The sensing loss is
where . Above the numerical floor, the logarithm preserves the ordering of the sample-average sensing objective while rescaling its gradient. Each block contributes a gradient through its precoder and the common geometry.
The nonnegative training margins , and tighten the rate and spacing requirements; evaluation uses the original thresholds. We define . The signed residuals are defined as , , and . Here, is evaluated using the repaired geometry and sample- precoder. Nonpositive residuals satisfy the corresponding tightened requirements. The rate residuals have units of bit/s/Hz, whereas is dimensionless. For a residual , multiplier , and coefficient , we use the one-sided Powell-Hestenes-Rockafellar (PHR) term [40]
Let contain the indices of the largest , and let contain the largest , where . With user-specific multipliers and pair-specific multipliers , all nonnegative, the constraint losses are
The positive coefficients control the augmented terms. Tail selection concentrates training on poorly served blocks and closely spaced raw coordinates. The joint primal loss is minimized over using Adam with decoupled weight decay (AdamW) with zero weight decay [41]. After each primal update, define and . The multipliers undergo projected signed-residual ascent:
for , every , and every . Here, are step sizes, are multiplier caps, and clips its argument to . All pair multipliers are updated, including pairs outside the current spacing tail. Signed residuals allow multipliers to decrease when their constraints are met with margin. The raw-spacing loss encourages separation before repair and the structured decoders enforce the movement region and transmit-power constraints directly.
Simulation Results
In this section, we evaluate the proposed TT-MBGNN in terms of angular sensing-error bounds, communication-QoS satisfaction, antenna-spacing feasibility, and online computational cost. All methods use the same test locations, underlying channel draws, reflection coefficients, and target-angle samples, with channels computed for each method’s geometry. The reported root-BCRB is degrees, where is the number of test blocks and is the target-averaged angular BCRB in radians squared.
Simulation Setup and Implementation
We consider a monostatic ISAC BS at the origin, equipped with MAs within the square movement region on the local - plane. In the default configuration, the BS serves single-antenna UEs and senses point targets. We use a uniform rectangular array (URA) within as the reference geometry . In each block, we sample the UE and nominal target locations independently and uniformly with respect to volume within and , respectively. Channels and target-prior centers vary across blocks, while we hold the geometry fixed within each slow interval. Table I summarizes the default principal parameters. With unity antenna-element and UE receive gains and no additional propagation losses, we compute the communication power gain as and the target reflection power as . Here, and are distances from the BS reference point, and we set for every target. We generate the complex reflection coefficient as , where is sampled independently across targets and blocks. We regenerate these coefficients across blocks and hold them fixed within each block, including across all angular-prior samples. We account for array gains through the steering vectors and precoders. In the physics engine, we use to generate channels and to evaluate the sensing FIM. For the BCRB, we treat the real and imaginary parts of as unknown nuisance parameters.
For each QoS pair, we jointly train the slow and fast policies using uniform per-UE requirements. We estimate the slow statistics from independent samples and hold these statistics fixed during training. At each training step, we draw independent block states and use angular-prior samples per block. We arrange the angular samples in symmetric pairs by accompanying each sampled offset with its negative about the prior center. We select checkpoints using a fixed validation set independent of the training and final test samples. For final evaluation, we use independent trials, each representing one slow interval of coherence blocks, giving unseen blocks per deterministic geometry or policy. Under unchanged statistical inputs, we reuse the geometry across trials and adapt the precoder to each block. We use angular-prior samples per test block, generated by the same paired procedure. We compute 95% confidence intervals from trial-level averages and apply first-order propagation through the square root for root-BCRB.
Benchmarks and Ablation Studies
We compare two learned-component ablations and three independent baselines on common test realizations under the same physical and communication requirements.
Learned-Component Ablations
Learned-Component Ablations: We use the following variants without retraining their learned components:
Fixed MA+F-MBGNN: We replace the learned positions with the feasible reference geometry to assess the contribution of geometry adaptation.
S-MBGNN+RZF: In this baseline, we combine the learned geometry with blockwise RZF to assess the contribution of learned precoding.
Independent Baselines
Fixed MA+RZF: We combine the reference geometry with blockwise RZF as a nonlinear benchmark.
SAA-AO: We approximate the long-term objective using finite sample averages and alternately update the common geometry and sample-dependent precoders from multiple initializations [10]. We also perform iterative precoder refinement on test blocks. The optimizer uses simulated target reflection coefficients to evaluate the sensing objective throughout optimization and refinement. In the proposed framework, we use these nuisance values only for offline loss computation and performance evaluation. Thus, SAA-AO serves as an idealized benchmark with additional sensing information and iterative computation at deployment.
Fixed MA+sensing-only eigenbeams: For the reference geometry, we use the leading eigenvectors of the steering-vector covariance averaged over samples from the current target-angle priors and normalize the precoder to the power budget. We include this sensing-oriented diagnostic without communication QoS constraints or direct angular BCRB minimization.
Training Convergence
Figure 3 presents the fixed-validation evolution under the six QoS configurations. The displayed curves use a ten-epoch trailing average for readability, whereas feasibility checks and final evaluation use unsmoothed metrics. In fig:3a, the root-BCRB decreases rapidly and then stabilizes within a QoS-dependent range. Residual fluctuations and the lack of strict ordering across independently trained models reflect stochastic sampling and nonconvex primal-dual optimization. More stringent communication requirements generally yield larger sensing bounds. fig:3b shows that the UE-averaged

Figure 3. Fixed-validation convergence of the proposed TT-MBGNN under uniform per-UE QoS pairs . In panel (b), solid curves show the UE-averaged 5th-percentile block rates, and dotted horizontal lines denote the fast-timescale requirements. In panel (c), solid and dashed curves show repaired and raw minimum spacings, respectively. The inset highlights several repaired curves approaching , with staggered markers distinguishing their overlap. The physical spacing requirement is . For metric curves, we use a ten-epoch trailing average for better visualization.
5th-percentile block rate settles above the corresponding fast-timescale requirement as the initially large RZF-based rate margin decreases and sensing performance improves. This aggregate percentile does not establish strict per-block feasibility. The maximum per-UE slow-rate shortfall was zero at every recorded validation epoch and is omitted. fig:3c shows that the raw minimum spacing can fall below , while the repaired curves remain close to . This observed behavior demonstrates the empirical effectiveness of the repair without establishing a universal spacing guarantee. Physical feasibility is assessed from the unsmoothed repaired coordinates against . The raw-spacing augmented-Lagrangian term separately discourages overlapping decoder outputs.
Sensing-Communication Tradeoff and qos Reliability
Figure 4 reports results on unseen test blocks. The fast-rate threshold varies from 1 to 6 bit/s/Hz, and the slow-rate threshold is 1 bit/s/Hz higher. The middle panel reports the fraction of blocks with a fast-rate violation by any UE defined as . The bottom panel shows , where is the test-mean rate of UE . These empirical checks are applied to each method and threshold pair. A sensing-bound value alone does not establish QoS feasibility and should be interpreted alongside the communication results. The proposed TT-MBGNN achieves a lower root-BCRB than both learned-component ablations throughout the sweep, supporting joint geometry and precoder adaptation. We observe positive slow-rate margins and low, nonzero fast outage for the proposed method. The remaining outages indicate that some test blocks violate the fast-rate requirements. The analytical RZF designs show no observed fast outage and larger slow-rate margins, together with higher sensing bounds. The sensing-only eigenbeams exhibit 98–100% fast outage and negative slow-rate margins. We omit their outage curve from the middle panel to improve the visibility of the other methods.

Figure 4. Root-BCRB (top), any-UE fast-QoS outage (middle), and minimum per-UE slow-rate margin (bottom) versus the uniform fast-QoS threshold. The slow threshold is bit/s/Hz. The sensing-only design is omitted from the middle panel because its fast outage exceeds 98% at every threshold. The dotted horizontal line denotes zero slow-rate margin.

Figure 5. Network-size transfer from with frozen parameters or ten fine-tuning epochs. Top: root-BCRB versus () and (). Bottom (): test-block mean worst-UE fast-rate shortfall, (solid, left axis) and any-UE fast-QoS outage (dashed, right axis) in the left panel, and minimum slow-rate margin in the right panel. Dotted lines mark the training size and zero. Fixed denotes the reference array. Sensing-only beams appear only in the top row.
Generalization Across Network Sizes
Fig. 5 evaluates policies trained with and uniform fast and slow QoS requirements of 3 and 4 bit/s/Hz, respectively. Additional-node centers and region radii are specified separately in Table I. The top panels show the root-BCRB as or varies. The bottom-left panel reports the test-block mean worst-UE fast-rate shortfall, , on the left axis and the any-UE fast-QoS outage percentage on the right axis. The latter is the percentage of test blocks in which at least one UE misses the fast-rate requirement. The bottom-right panel reports the minimum slow-rate margin, .
The frozen TT-MBGNN achieves favorable sensing bounds across the tested numbers of UEs and targets. Increasing beyond the training size raises the communication load, resulting in larger fast-rate shortfalls, more frequent outages, and eventually a negative slow-rate margin. Ten fine-tuning epochs, corresponding to 5% of the original training epoch count, substantially reduce fast-rate shortfalls and outages at these larger user counts while maintaining favorable sensing bounds. The fine-tuned policies also yield positive slow-rate margins across all tested user counts. The remaining nonzero fast outage indicates residual block-level violations. These results support reuse of the trained policies with limited adaptation to increased communication load.
Online Computational Cost
Table II reports batch-one precoder-update latency at , measured on an NVIDIA GeForce RTX 5080 GPU (16 GB) with an Intel Core Ultra 9 275HX
| Method | Proposed TT-MBGNN | SAA-AO (100 updates) |
| Median (ms) | 16.48 | 832.80 |
Table II. Median Batch-One Fast-Update Latency
CPU. The proposed TT-MBGNN requires 16.48 ms per block, compared with 832.80 ms for SAA-AO with 100 updates, giving a 50.5-fold speedup. These timings exclude channel acquisition and antenna movement. The speedup depends on network size, hardware, and iteration budget, while the proposed fast policy requires no online optimization loop.
Conclusion
In this paper, we developed a two-timescale MA-assisted ISAC framework with block-level and long-term per-UE rate requirements. The proposed TT-MBGNN combines an S-MBGNN for antenna positioning with an F-MBGNN for block-adaptive precoding. Relation-aware layers represent interactions among antennas, users, and targets. A bounded position decoder is followed by finite spacing repair, with a STE supplying surrogate gradients during training. The fast decoder learns complex corrections around an RZF reference. Joint training combines a Monte Carlo angular BCRB objective with one-sided augmented-Lagrangian terms for rate requirements and raw-spacing violations. Simulations demonstrated favorable sensing-communication tradeoffs relative to the principal ablations and analytical baselines. The proposed method achieved competitive sensing bounds against the idealized SAA-AO benchmark and approximately 50-fold faster batch-one precoder computation under the reported settings. Frozen policies provided competitive sensing bounds at the tested unseen user and target counts, while brief fine-tuning improved communication reliability at larger user counts. Communication reliability was assessed empirically, with residual fast-rate outages. Every repaired geometry checked in the reported simulations met the minimum antenna spacing requirement, although the finite repair provides no universal spacing guarantee. Future work will examine robustness to imperfect CSI and array calibration errors and validate the framework on practical MA platforms.
Appendix A: Input-Feature Construction
This appendix gives the explicit construction and normalization of the feature collections in (32) and (33). Each node-feature vector below is a column vector whose transpose forms the corresponding row of its feature matrix. For example, . The same convention applies to the remaining slow- and fast-timescale features. The vectors are columns of and , respectively.
Slow-Timescale Features
The reference position is normalized to the movement region as
For UE and , let be the corresponding line-of-sight departure angle and define its circular mean by
Let denote the large-scale channel power gain and the Rician factor in linear units. Their scaled statistics are and , where is a fixed offset and is a scale. With rate scale in bit/s/Hz, the feature is
The expectations are computed offline from the long-term distributions using (13). For target , define the nominal angle , where is the Cartesian-to-angle map used in the system model. The angular widths are . The periodic feature
avoids discontinuities from angular wrapping [42].
Fast-Timescale Features
For the fast MA feature, we use the selected geometry:
This vector is constant within one positioning interval but generally differs from the reference feature in (50). For the target’s features, we form the following using the prior:
For channel coefficient , define and the real-valued compression . The channel feature is
The tensor entry stores this vector. Both directed computational edges for pair use the same feature. Reversing a graph edge does not conjugate the physical channel coefficient. With , the UE feature is
The node-level channel summaries complement the link-level complex channel information in . Together, (54), (55), (56), and (57) provide the entries of defined in (33).
Appendix B: Relation-Aware Layer Implementation
In (36), the score matrix has no bias. Each -dimensional message is divided into consecutive heads of dimension . Attention weights are normalized over the source nodes of the same relation and destination. During training, dropout is applied to these weights before the attention aggregate in (37) is formed; the mean aggregate is computed from the original messages. Attention dropout is disabled during inference. The affine self transformation and gate map used in (38) and (39) are given by , . Here, , , and both bias vectors belong to . The gate is . The message function , relation fusion function , and type-specific fusion function each use two affine layers with a -dimensional hidden layer, a SiLU activation, and hidden-layer dropout. Relations entering a node are concatenated in a fixed order. The MA and UE nodes each receive two relation types, whereas target nodes receive one. Parameters are shared among nodes of the same type within a policy, but are distinct across layers and between the slow and fast policies.
Pairwise RZF-Residual Precoder Decoder
Normalize the reference coefficient as . Using the compression defined in Appendix A, form
The shared MLP produces the real correction vector :
Scaling and normalization then follow (45).
References
- [1]A. Liu et al., “A survey on fundamental limits of integrated sensing and communication,” IEEE Commun. Surveys Tuts., vol. 24, no. 2, pp. 994–1034, 2022.arxiv.org/abs/2104.09954
- [2]S. Lu et al., “Integrated sensing and communications: Recent advances and ten open challenges,” IEEE Internet Things J., vol. 11, no. 11, pp. 19094–19 120, 2024.arxiv.org/abs/2305.00179
- [3]L. Zhu et al., “Movable antennas for wireless communication: Opportunities and challenges,” IEEE Commun. Mag., vol. 62, no. 6, pp. 114–120, 2024.arxiv.org/abs/2306.02331
- [4]W. Ma et al., “A survey on reconfigurable and movable antennas for wireless communications and sensing,” IEEE Commun. Surveys Tuts., vol. 28, pp. 4842–4882, 2026.DOI
- [6]A. Fadakar et al., “Hybrid codebook design for localization using electromagnetically reconfigurable fluid antenna system,” IEEE J. Sel. Topics Signal Process., pp. 1–16, 2026.
- [7]Y. Zhang et al., “Movable antenna-aided hybrid beamforming for multi-user communications,” IEEE Trans. Veh. Technol., vol. 74, no. 6, pp. 9899–9903, 2025.arxiv.org/abs/2404.00953
- [8]C. Jiang et al., “Movable antenna-assisted integrated sensing and communication systems,” IEEE Trans. Wireless Commun., vol. 24, no. 8, pp. 6397–6412, 2025.arxiv.org/abs/2501.01217
- [9]L. Chen et al., “Antenna position and beamforming optimization for movable antenna enabled ISAC: Optimal solutions and efficient algorithms,” IEEE Trans. Signal Process., vol. 73, pp. 3812–3828, 2025.arxiv.org/abs/2502.14198
- [10]W. Ma et al., “Movable antenna enhanced integrated sensing and communication via antenna position optimization,” IEEE Trans. Signal Process., vol. 74, pp. 1522–1537, 2026.arxiv.org/abs/2501.07318
- [11]Y. Han et al., “Movable antenna enhanced downlink multi-user integrated sensing and communication system,” in Proc. IEEE Veh. Technol. Conf. (VTC Spring), 2025, pp. 1–6.arxiv.org/abs/2503.22486
- [12]H. Qin et al., “Cramér-Rao bound minimization for movable antenna-assisted multiuser integrated sensing and communications,” IEEE Wireless Commun. Lett., vol. 13, no. 12, pp. 3404–3408, 2024.arxiv.org/abs/2407.11875
- [13]A. Khalili et al., “Movable antenna-enabled ISAC: Tackling slow antenna movement, dynamic RCS, and imperfect CSI via two-timescale optimization,” IEEE Trans. Wireless Commun., vol. 25, pp. 8452–8467, 2026.DOI
- [14]Z. Zheng et al., “Two-timescale design for movable antenna-enabled multiuser MIMO systems,” IEEE Trans. Commun., vol. 73, no. 11, pp. 10 554–10 571, 2025.DOI
- [15]R. D. Gill et al., “Applications of the van trees inequality: A bayesian Cramér-Rao bound,” Bernoulli, vol. 1, no. 1/2, pp. 59–79, 1995.DOI
- [16]Y. Zhang et al., “6D movable antenna enhanced cell-free MIMO: Two-timescale decentralized beamforming and antenna movement optimization,” arXiv preprint arXiv:2601.04969, 2026.arxiv.org/abs/2601.04969
- [17]Y. Shen et al., “Graph neural networks for wireless communications: From theory to practice,” IEEE Trans. Wireless Commun., vol. 22, no. 5, pp. 3554–3569, 2023.arxiv.org/abs/2203.10800
- [18]Y. Lu et al., “Graph neural networks for wireless networks: Graph representation, architecture and evaluation,” IEEE Wireless Commun., vol. 32, no. 1, pp. 150–156, 2025.arxiv.org/abs/2404.11858
- [19]Y. Li et al., “GNN-based beamforming for sum-rate maximization in MU-MISO networks,” IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 9251–9264, 2024.arxiv.org/abs/2311.03659
- [20]R. Zhang et al., “Model-based GNN enabled energy-efficient beamforming for ultra-dense wireless networks,” IEEE Trans. Wireless Commun., vol. 24, no. 4, pp. 3333–3345, 2025.arxiv.org/abs/2410.02289
- [21]Z. Wang et al., “Heterogeneous graph neural network for cooperative ISAC beamforming in cell-free MIMO systems,” in Proc. ACM Int. Conf. Mobile Comput. Netw. (MobiCom), 2024, pp. 2161–2172.arxiv.org/abs/2410.09963
- [23]U. Demirhan et al., “Learning beamforming in cell-free massive MIMO ISAC systems,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2024, pp. 326–330.arxiv.org/abs/2409.18237
- [24]C. Weng et al., “Learning-based joint beamforming and antenna movement design for movable antenna systems,” IEEE Wireless Commun. Lett., vol. 13, no. 8, pp. 2120–2124, 2024.arxiv.org/abs/2404.01784
- [27]M. Sheraz et al., “RSMA-assisted fluid antenna ISAC via hierarchical deep reinforcement learning,” in Telecom, vol. 7, no. 2. MDPI, 2026, p. 41.DOI
- [28]C. He et al., “Graph neural network enabled fluid antenna systems: A two-stage approach,” IEEE Trans. Veh. Technol., vol. 74, no. 10, pp. 16 625–16 629, 2025.arxiv.org/abs/2502.03922
- [29]Y. Xiu et al., “Distributed distortion-aware robust optimization for movable antenna-aided cell-free ISAC systems,” arXiv preprint arXiv:2508.13839, 2025.
- [30]F. Liang et al., “Two-timescale deep optimization of positioning and beamforming in movable antenna arrays,” IEEE Trans. Wireless Commun., vol. 25, pp. 14 242–14 258, 2026.DOI
- [31]L. Zhu et al., “Modeling and performance analysis for movable antenna enabled wireless communications,” IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6234–6250, 2024.arxiv.org/abs/2210.05325
- [32]S. Bay et al., “On the hybrid Cramér Rao bound and its application to dynamical phase estimation,” IEEE Signal Process. Lett., vol. 15, pp. 453–456, 2008.
- [33]A. Fadakar et al., “Mutual coupling-aware localization for RIS-assisted ISAC systems,” IEEE Trans. Cogn. Commun. Netw., vol. 11, no. 5, pp. 2938–2954, 2025.DOI
- [34]H. L. Van Trees, Detection, estimation, and modulation theory, part III: radar-sonar signal processing and Gaussian signals in noise. John Wiley & Sons, 2001.
- [35]X. Xiao et al., “Multiuser communications with movable-antenna base station: Joint antenna positioning, receive combining, and power control,” IEEE Trans. Wireless Commun., vol. 23, no. 12, pp. 19 744–19 759, 2024.arxiv.org/abs/2308.09512
- [36]M. Eisen et al., “Optimal wireless resource allocation with random edge graph neural networks,” IEEE Trans. Signal Process., vol. 68, pp. 2977–2991, 2020.DOI
- [37]M. Schlichtkrull et al., “Modeling relational data with graph convolutional networks,” in Proc. Eur. Semantic Web Conf. (ESWC). Springer, 2018, pp. 593–607.arxiv.org/abs/1703.06103
- [38]P. Veličković et al., “Graph attention networks,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2018.arxiv.org/abs/1710.10903
- [39]Y. Bengio et al., “Estimating or propagating gradients through stochastic neurons for conditional computation,” arXiv preprint arXiv:1308.3432, 2013.arxiv.org/abs/1308.3432
- [40]R. T. Rockafellar, “Augmented lagrange multiplier functions and duality in nonconvex programming,” SIAM Journal on Control, vol. 12, no. 2, pp. 268–285, 1974.DOI
- [41]I. Loshchilov et al., “Decoupled weight decay regularization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2019.
- [42]K. V. Mardia et al., Directional statistics. John Wiley & Sons, 2009.

