Intelligent Base Station Deployment in Urban Wireless Networks: A Geographic Data-Informed Digital Twin Approach

Zhenyu Tao; Yuxuan Li; Wei Xu; Yongming Huang; Xiaohu You

Source-linked AI summary

Intelligent Base Station Deployment in Urban Wireless Networks: A Geographic Data-Informed Digital Twin Approach

Zhenyu Tao, Yuxuan Li, Wei Xu, Yongming Huang, Xiaohu You

arXiv:2608.14599v1Jun 30, 2026cs.NIcs.AI

TL;DR

Urban base-station deployment requires difficult-to-obtain radio and user-distribution information before deployment. This paper combines a geographic-data-informed digital twin with deep reinforcement learning, achieving up to 98.9% of idealized deployment performance while reducing optimization overhead by more than 99%.

Problem
Urban base-station deployment requires site-specific radio maps and realistic user distributions that are difficult to obtain before deployment, while optimization is combinatorial.
Method
The framework combines an open-geographic-data wireless digital twin with deep reinforcement learning for sample-free macro-base-station deployment optimization.
Results
Up to 98.9% of idealized deployment performance was achieved while optimization overhead was reduced by more than 99%.
Takeaways & Limitations
Geographic-data-informed digital twins can support near-ideal urban macro-base-station deployment with substantially lower optimization overhead.

Abstract

from arXiv · show

The placement of base station (BS) is a fundamental determinant of coverage and capacity of urban wireless networks. Yet large-scale BS deployment optimization remains challenging due to its dependency on site-specific radio propagation and user spatial distributions, both of which are unfortunately difficult to obtain prior to deployment. To overcome this barrier, we propose an intelligent BS deployment framework that integrates a geographic data-informed wireless network digital twin (DT) with deep reinforcement learning (DRL), enabling sample-free macro BS deployment optimization from solely open geographic data, without on-site measurements, real user trajectories, or exhaustive ray tracing. The proposed DT incorporates a sample-free radio map prediction model with hybrid input representation to achieve kilometer-scale signal strength estimation in milliseconds, complemented by a diffusion-based generative model for trajectory synthesis to collectively characterize channel and user distributions. Leveraging the DT as a virtual training environment, we formulate BS deployment as a multi-step Markov decision process (MDP) and solve it via a spatially structured DRL algorithm. A local search process and a Wasserstein distance-based deployment buffer are further incorporated to efficiently explore the large combinatorial solution space. Experimental results in real-world urban scenarios demonstrate that the geographic data-informed DT attains accuracy comparable to 100-sample-based prediction, and the intelligent BS deployment framework achieves up to 98.9% of the idealized benchmark performance while reducing optimization overhead by over 99%.

I. INTRODUCTION … C. User Mobility Modeling

The paper presents a geographic data-informed wireless-network digital twin integrated with deep reinforcement learning for sample-free urban macro base-station deployment. It combines rapid radio-map prediction and synthetic mobility modeling to support capacity-aware optimization, achieving up to 98.9% of idealized performance with over 99% lower optimization overhead.

I. INTRODUCTION: Urban base-station placement fundamentally constrains coverage, capacity, and the effectiveness of post-deployment network optimization, especially in dense areas with complex building landscapes.Post-deployment methods include resource scheduling, beamforming, and user association, but deployment locations determine achievable network performance.
I. INTRODUCTION: Existing deployment methods rely on statistical channel models or prior channel-gain assumptions, which can produce suboptimal decisions in complex urban propagation environments.The literature spans heuristic, stochastic-geometry, convex-optimization, and DRL approaches.
B. Radio Map Construction: Ray tracing provides high-fidelity deterministic channels, but its computational cost motivates neural-network radio-map methods that avoid expensive deployment-time measurements.Sample-based methods use sparse on-site measurements, creating a circular dependency because measurements are difficult to obtain before deployment.
C. User Mobility Modeling: User spatial distribution is coupled with channel conditions and jointly affects link quality, cell load, bandwidth allocation, and capacity evaluation.Separately assuming uniform users or stochastic mobility can ignore urban street-topology constraints and misrepresent rush-hour traffic demand.
I. INTRODUCTION: The proposed framework uses openly available OpenStreetMap data to construct a wireless-network digital twin for sample-free macro base-station deployment without measurements, real trajectories, or exhaustive ray tracing.The framework integrates the geographic-data-informed digital twin with deep reinforcement learning.
I. INTRODUCTION: The digital twin predicts kilometer-scale radio maps in milliseconds using hybrid inputs for 3D building heights and material-specific electromagnetic properties, while synthesizing trajectories from hierarchical street maps.Together, these models provide coupled estimates of user distribution and channel conditions for a given urban area.
I. INTRODUCTION: The deployment task is formulated as a multi-step Markov decision process and solved with spatially structured deep reinforcement learning, local search, and a Wasserstein-distance-based deployment buffer.These components improve exploration of the combinatorial deployment space, refine solutions, preserve diversity, and mitigate local-optimum convergence.
I. INTRODUCTION: 98.9% of the idealized deployment performance is achieved while optimization overhead is reduced by more than 99%.Validation uses real geographic data, authentic user trajectories, and ray-tracing results; ablations show hybrid inputs achieve accuracy comparable to sample-based methods.

III. PROBLEM FORMULATION … 2) Capacity Objective:

The section formulates base-station deployment as a multiobjective optimization problem jointly considering outdoor-area coverage and cumulative network capacity. It models capacity through max-RSS cell association, SINR, bandwidth sharing, spectral efficiency, and per-user throughput across spatially distributed users.

III. PROBLEM FORMULATION: The deployment problem jointly optimizes network coverage and system capacity.Coverage and capacity are treated as the two objectives of the BS deployment task.
A. System Model: The urban area A is discretized into Np × Np pixels, with outdoor pixels O representing streets and open spaces where users may be located.Buildings with varied heights and materials are modeled as radio-propagation obstacles.
A. System Model: The framework selects M BSs from K rooftop candidates, where each candidate is defined by horizontal coordinates and antenna height.The selected subset B ⊆ Z satisfies |B| = M and M ≪ K, with the goal of maximizing overall network performance.
1) Coverage Objective:: Coverage measures the fraction of outdoor area receiving adequate signal strength under max-RSS association.RSS is computed from transmit power and path loss, and adequacy is evaluated against the minimum sensitivity threshold ψthr.
2) Capacity Objective:: Each outdoor pixel is associated with the BS providing the strongest signal, partitioning O into M disjoint serving regions.The serving BS is denoted cij, and the corresponding regions are represented by ΩB_m.
2) Capacity Objective:: Capacity is derived from per-pixel SINR and user spatial distribution, incorporating interference, thermal noise, equal-share bandwidth, and capped spectral efficiency.The network capacity is defined as cumulative throughput across all users, with spectral efficiency bounded by ηmax.

C. Optimization Problem

The section formulates BS deployment as maximizing a weighted combination of coverage and expected system capacity under a fixed-size candidate subset constraint. It also identifies unavailable site-specific radio maps and user densities, combinatorial search, and nonconvex propagation effects as key optimization challenges.

Objective formulation: Expected throughput depends on the expected normalized user distribution ρ(i, j), with randomness arising from dynamic user mobility while SINR remains deterministic for a given deployment.The expected user distribution can be estimated from real-world or synthetically generated trajectory datasets.
Objective formulation: The optimal deployment B∗ maximizes a weighted combination of coverage and expected system capacity, subject to selecting exactly M BSs from candidate set Z.The weighting factor β balances coverage and capacity, while Rnorm normalizes throughput to a comparable range.
Optimization challenges: Prior to deployment, the objective requires site-specific radio maps ψ and realistic user spatial density ρ, which are difficult to obtain.These unavailable inputs motivate the need for a geographic data-informed optimization framework.
Optimization challenges: K could exceed 10^4 in kilometer-scale urban areas, making selection of M BSs a large combinatorial subset-selection problem.The resulting solution space is too large for straightforward exhaustive optimization.
Optimization challenges: Complex propagation, cell association dynamics, and inter-cell interference create nonconvex dependence on BS locations, precluding usual gradient-based methods.Both coverage and capacity exhibit this nonconvex dependence.

IV. GEOGRAPHIC DATA-INFORMED WIRELESS NETWORK DIGITAL TWIN … 2) Hybrid Input Representation:

The proposed geographic data-informed wireless-network digital twin enables sample-free macro base-station deployment optimization from open geographic data. It combines kilometer-scale radio-map prediction with generative user-mobility modeling and a hybrid representation of urban geometry, transmitter settings, and electromagnetic properties.

IV. GEOGRAPHIC DATA-INFORMED WIRELESS NETWORK DIGITAL TWIN: The digital twin estimates radio maps and user distributions directly from open geographic data to support intelligent base-station deployment.It combines a sample-free radio-map prediction model with a generative user-mobility model.
A. Sample-free Radio Map Prediction Model: Sample-free radio-map prediction replaces costly field measurements or on-demand ray tracing by inferring site-specific propagation from environmental descriptions.Existing approaches typically operate on compact 256 × 256 m2 regions, limiting scalability across heterogeneous urban environments.
A. Sample-free Radio Map Prediction Model: A kilometer-scale dataset incorporates 3D urban morphology and material-specific electromagnetic properties tailored to macro base-station deployment.The prediction model estimates site-specific radio maps directly from geographic data.
1) Radio Map Dataset Construction:: Incomplete building heights are completed with a density-aware stochastic generation method using category-specific log-normal sampling and physically plausible clipping.Generated heights are clipped between 6 m and 300 m.
1) Radio Map Dataset Construction:: Buildings receive heterogeneous material types based on ITU-R P.2040, with four candidate building materials and a ground material characterized by permittivity and conductivity.Material assignments model realistic shared construction properties within commercial complexes or residential compounds.
1) Radio Map Dataset Construction:: Sionna generates site-specific radio maps through ray tracing from the processed geographic data.These maps provide the radio-map data used in constructing the kilometer-scale dataset.
2) Hybrid Input Representation:: The hybrid input X ∈R4×Np×Np encodes 3D building heights, base-station configuration, conductivity, and relative permittivity, producing a single-channel RSS map Y ∈R1×Np×Np.The representation captures diffraction, reflections, geometric obstruction, transmitter configuration, and material-dependent electromagnetic interactions, while remaining extensible to sample-based prediction.

3) Network Architecture and Data Augmentation: … 2) Estimation of User Spatial Distribution:

The framework combines an architecture-agnostic, augmented radio-map predictor with Map2Traj, a diffusion-based zero-shot mobility model conditioned on OpenStreetMap street maps. Together, these models form a geographic-data-driven digital twin that jointly estimates channel conditions and user spatial distributions without real trajectory data, enabling millisecond-level evaluation of deployment candidates.

3) Network Architecture and Data Augmentation:: The hybrid input representation integrates with U-Net, PMNet, and ViT-based RMTransformer backbones.Training randomly applies one of eight geometric transformations: four 90° rotations and their horizontal mirrors.
3) Network Architecture and Data Augmentation:: Millisecond-level radio-map prediction is available for any candidate BS location using only openly available geographic data.The trained model predicts the radio map ˆψ without on-site measurements or real user trajectories.
B. Generative User Mobility Model: Map2Traj eliminates the need for area-specific real trajectory datasets by synthesizing user mobility from hierarchical OpenStreetMap street-map images alone.This zero-shot model is built upon a denoising diffusion probabilistic model and estimates ˆρ for arbitrary target areas.
1) Diffusion-based Trajectory Generation:: Map2Traj encodes hierarchical road types and trajectories as multi-channel and single-channel images, respectively.Road categories such as Trunk, Primary, and Residential are represented as separate binary channels.
1) Diffusion-based Trajectory Generation:: The DDPM-based generator progressively adds Gaussian noise to trajectories and learns reverse denoising conditioned on the street map.Continuous street-map guidance preserves road-topology constraints, including in areas unseen during training.
2) Estimation of User Spatial Distribution:: Synthetic trajectories are aggregated and normalized to estimate the user spatial distribution ˆρ from the target area’s OpenStreetMap street map.The estimation uses neither real user trajectory data nor privacy-sensitive information.
2) Estimation of User Spatial Distribution:: The radio-map and mobility models use the same geographic source to coherently couple building-governed propagation with street-constrained user mobility.The resulting digital twin provides ˆψ and ˆρ for evaluating the coverage-capacity objective at the millisecond level for any deployment candidate B.

V. DRL-BASED INTELLIGENT BS DEPLOYMENT VIA WIRELESS NETWORK DIGITAL TWIN · A. MDP Formulation · 1) State Space:

The paper formulates macro BS deployment as a sequential multi-step MDP evaluated by a wireless-network digital twin and solves it with spatially structured SAC enhanced by local search and a Wasserstein deployment buffer. Each state combines candidate availability, deployed-BS positions, and composite radio coverage to guide spatially complementary placements and identify coverage gaps.

V. DRL-BASED INTELLIGENT BS DEPLOYMENT VIA WIRELESS NETWORK DIGITAL TWIN: The framework uses the wireless-network digital twin to optimize the combinatorial BS deployment problem through sequential decision-making.It evaluates network performance after each BS placement.
V. DRL-BASED INTELLIGENT BS DEPLOYMENT VIA WIRELESS NETWORK DIGITAL TWIN: The proposed solver is a spatially structured soft actor-critic algorithm enhanced with local search and a Wasserstein distance-based deployment buffer.These components are introduced to address the combinatorial deployment optimization problem.
A. MDP Formulation: The deployment process is modeled as a multi-step MDP that selects M BS locations from candidate set Z.The MDP tuple is (S, A, R, P, γ), representing states, actions, rewards, transitions, and discounting.
1) State Space:: At step n, the state sn consists of three spatial feature maps, each sized Np × Np.Together, the maps represent available sites, deployed BSs, and current signal coverage.
1) State Space:: The available-location map is initialized with all K candidate positions in Z and removes candidates within inhibition radius rinh after deploying a BS at zm.This enforces a minimum inter-BS separation.
1) State Space:: The deployed-BS map records positions selected before step n, enabling the policy to reason about spatial complementarity with existing BSs.It is defined by Sn^d(i, j) = I ((i, j) ∈ Bn−1).
1) State Space:: The coverage map uses the pixel-wise maximum RSS from previously deployed BSs, allowing the policy to identify coverage gaps for subsequent placements.The previously deployed set satisfies Bn−1 ⊆ Z.
1) State Space:: At step n, the action selects a single pixel (x, y) from the Np × Np grid as the next BS location.This defines each deployment decision within the spatial state representation.

2) Action Space: … 3) Actor Update:

The framework restricts BS-placement actions to valid candidates, uses incremental DT-based rewards with deterministic state transitions, and applies a spatially structured discrete SAC algorithm. U-Net actor–critic networks produce pixel-level deployment decisions, with critic and actor updates designed for stable, entropy-aware exploration.

2) Action Space:: The candidate BS map Sc filters the policy output so that only valid deployment positions can be selected.The action space is constrained by the candidate BS map.
3) Reward Function:: The reward is the incremental objective gain from each newly placed BS, with R1 defined directly and later rewards computed as Rn = R(Bn|ˆψ, ˆρ) − R(Bn−1|ˆψ, ˆρ).The episode terminates immediately after the M-th BS is deployed.
4) State Transition:: State transitions are deterministic: the deployed and candidate maps are updated, while the wireless network DT recomputes and merges the new BS signal-strength map.The candidate map follows the inhibition rule in (19), and radio-map merging follows (20).
B. Spatially-Structured Soft Actor-Critic Algorithm: The algorithm adopts discrete SAC, maximizing an entropy-regularized expected return to balance exploration of the combinatorial deployment space with training stability.Entropy regularization supports exploration while preserving stable training.
1) Network Architecture:: Actor and critic networks use residual U-Net backbones whose encoders capture multi-scale spatial features and decoders restore Np×Np resolution through skip connections.This architecture preserves fine-grained spatial details for pixel-level BS deployment.
1) Network Architecture:: The actor outputs a single-channel RNp×Np probability map, while two critics output pixelwise Q-value maps and target critics use Polyak averaging for stability.Each pixel represents the estimated value of placing a BS at that location.
2) Critic Update:: The target soft value averages over valid next-state actions with entropy regularization, using a clipped double-Q estimate formed from the two target critics.The clipped estimate is min(s, a) = mink∈{1,2} Q¯ϕk(s, a), with α as the entropy temperature.
3) Actor Update:: Critics minimize MSE against the target Q-value, whereas the actor minimizes a discrete SAC KL-divergence objective that favors high-Q locations while retaining entropy.The actor uses Qmin(s, a) = mink∈{1,2} Qϕk(s, a), and α is automatically adjusted toward a target entropy.

4) Automatic Entropy Tuning: … D. Wasserstein Distance-based Deployment Buffer

The method adaptively tunes exploration during DRL training, then combines global policy exploration with millisecond-scale local search and a Wasserstein-diverse deployment buffer to refine multiple candidate deployments.

4) Automatic Entropy Tuning:: log α is optimized jointly with the actor and critic, automatically adapting exploration throughout each episode.¯H controls the exploration degree.
5) Training Procedure:: At each training episode, the agent sequentially places M BSs and stores transitions in replay buffer D before updating critics, actor, and entropy temperature.Target critics use Polyak averaging, and gradient clipping prevents training instability.
C. Local Search Refinement: The local search process refines DRL-generated deployments after training because stochastic exploration may not traverse all locations in Z.It complements DRL’s global exploration with fine-grained coordinate-level optimization.
C. Local Search Refinement: For each BS zm, local search constructs candidate grid Ωm within radius rref, evaluates modified deployments through the DT, and relocates zm to the highest-objective position.Each alternative uses Bzm←ω ≜B\{zm}∪{ω}, with the remaining M −1 BSs fixed.
C. Local Search Refinement: Within seconds, the entire local search process completes even for large candidate sets because the DT evaluates objectives at millisecond-level speed.Search repeats until no improvement is observed in a full round.
D. Wasserstein Distance-based Deployment Buffer: The quality of the LS-refined solution depends heavily on its starting deployment because training episodes are orders of magnitude fewer than the explored deployment possibilities.Consequently, globally optimal deployments may remain unvisited and LS from one deployment is limited to its local neighborhood.
D. Wasserstein Distance-based Deployment Buffer: The deployment buffer Q accumulates diverse training deployments to provide multiple post-training starting points, avoiding reliance on only the single best explored deployment.Retaining only top-Nbuf reward-ranked deployments can cluster candidates spatially and converge them to one local optimum.
D. Wasserstein Distance-based Deployment Buffer: Q stores at most Nbuf ordered deployments as tuples (B, R(B|ˆψ, ˆρ)), where B contains BS coordinates and R(B|ˆψ, ˆρ) is the reward from (11).The buffer is initialized as empty at the beginning of training, and Wasserstein distance enforces spatial diversity.

1) Buffer Structure: … 1) Dataset and Scenario:

The framework preserves diverse, high-reward BS deployments using a Wasserstein-distance buffer, then applies local search to refine buffered solutions. Experiments evaluate the approach with geographically grounded radio-map and mobility data from realistic Xi’an urban scenarios.

2) Wasserstein Distance-based Diversity Metric:: The deployment buffer uses Wasserstein distance to measure structural dissimilarity between unordered BS-coordinate sets, with larger values indicating more distinct configurations.The distance is computed as an optimal matching problem and solved by the Hungarian algorithm in O(M^3) time.
3) Diversity-Preserving Update Rule:: For nearby deployments with W < δ, the buffer retains the higher-reward deployment by replacing the existing neighbor only when R(Bnew) > R(B(q∗)).This rule preserves the superior candidate within each spatial neighborhood.
3) Diversity-Preserving Update Rule:: Structurally distinct deployments with W(Bnew, B(q∗)) ≥ δ are appended when capacity remains, or replace the lowest-reward entry when the full buffer benefits from the newcomer.A full buffer discards Bnew unless R(Bnew) exceeds the minimum reward currently stored.
4) Post-Training Buffer Exploitation:: After DRL training, local search independently refines every buffered deployment, and the final solution maximizes the post-refinement objective R(LS(B)).The three-stage pipeline combines global DRL exploration, diversity preservation, and fine-grained local refinement.
VI. EXPERIMENT RESULTS: The experiments assess geographic-data-informed radio-map prediction and DRL-based BS deployment optimization under realistic urban scenarios constructed from open geographic data.The evaluation covers both the wireless-network digital twin and the deployment optimizer.
A. Experimental Settings: The Xi’an dataset uses OpenStreetMap building footprints and street networks, with 80 areas split into 64 training, 8 validation, and 8 test areas and randomized 30 times per area.Each materialized area includes 25 rooftop BS placements for scenario construction.
1) Dataset and Scenario:: 60,000 radio map samples are generated as 80 × 30 × 25 using 53 dBm transmission at 3.5 GHz and ray tracing with diffraction and maximum reflection depth 10.The dataset jointly captures building geometry, height variation, BS placement, and material diversity across over 60 km2.
1) Dataset and Scenario:: The evaluation uses Xi’an rush-hour vehicle trajectories with geographically separated Map2Traj training and test sets, while testing M ∈ {4, 5, 6} BSs across roughly 15,000 candidate rooftop locations.The coverage threshold is ψthr = −80 dBm, and β ∈ {0.5, 1} represents joint coverage-capacity and pure coverage objectives.

2) Model Setting and Training: · 3) Compared Methods:

The study trains and compares multiple radio-map backbones and input representations, then evaluates deployment strategies spanning geometry-based, ray-tracing, and digital-twin methods. The DRL deployment agent uses spatial SAC with U-Net actor and critic networks, while deployment variants add local search, buffering, and optional ray-tracing selection.

2) Model Setting and Training:: Three radio-map backbones are evaluated: five-level UNet, PMNet with atrous spatial pyramid pooling, and a ViT-based encoder-decoder.All models use MSE loss and the Section IV-A data augmentation technique.
2) Model Setting and Training:: The spatial SAC deployment agent uses U-Net backbones for both actor and critic networks, automatically tunes α, and updates target critics with Polyak averaging.The entropy temperature α is tuned via (25).
3) Compared Methods:: The 2D, 3D, and 3D+EM representations progressively include building footprints, heights, antenna-height encoding, and electromagnetic material properties.The 3D+EM configuration includes conductivity and permittivity maps.
3) Compared Methods:: X+Sample augments these configurations with 100 sparse measurement samples, while radio maps are evaluated using outdoor-pixel MSE and coverage prediction accuracy.The samples use mask and measured-signal-strength channels; CPA measures agreement between predicted and ground-truth coverage states at a signal-strength threshold.
3) Compared Methods:: Deployment comparisons include Hexagonal, genie-aided exhaustive RT, conventional exhaustive RT, DT-Greedy, and several DT-DRL variants.The genie-aided exhaustive RT method uses ground-truth ray-tracing maps and measured user distributions as the optimal benchmark, whereas conventional exhaustive RT assumes uniform users.
3) Compared Methods:: DT-DRL variants retain the best training deployment or refine it with LS, Wasserstein-buffer search, and optionally final ray-tracing evaluation.The variants are DT-DRL, DT-DRL+LS, DT-DRL+LS+Buffer, and DT-DRL+LS+Buffer+RT.

B. Radio Map Prediction Results · 1) Performance Comparison: · 2) Qualitative Analysis:

The 3D+EM hybrid representation consistently improves radio map prediction, with sample-free PMNet approaching sample-based accuracy. Qualitative results show that building geometry and electromagnetic properties capture height-dependent propagation and reflection-prone edge effects.

1) Performance Comparison:: PMNet’s sample-free MSE falls from 0.0203 with 2D input to 0.0135 with 3D and 0.0126 with 3D+EM.These correspond to 33.7% and 38.0% reductions relative to the 2D baseline.
1) Performance Comparison:: Sample-free PMNet’s CPA gap versus sample-based prediction narrows from 5.92% with 2D to 0.84% with 3D+EM.The sample-free 3D+EM setting achieves CPA 0.8836 versus 0.8920 for its sample-based counterpart.
1) Performance Comparison:: U-Net retains a substantial gap to sample-based prediction, whereas PMNet and ViT effectively exploit enriched 3D+EM inputs through multi-scale or global attention mechanisms.The results indicate that U-Net’s limited receptive field constrains its use of additional geometric and material information.
1) Performance Comparison:: The 3D+EM representation captures dominant propagation mechanisms and enables near-sample-based, sample-free prediction accuracy.PMNet is selected for subsequent radio map analysis and BS deployment optimization because it balances prediction accuracy and computational efficiency.
2) Qualitative Analysis:: The 2D model produces overly conservative coverage because it cannot distinguish low-rise buildings permitting over-rooftop propagation from tall signal-blocking structures.Adding building height enables the 3D model to capture vertical obstruction and over-rooftop propagation effects.
2) Qualitative Analysis:: 3D geometry captures height-dependent propagation effects, while electromagnetic properties further improve accuracy in reflection-prone edge regions.Their combined effect validates the hybrid input representation for urban macro BS radio map prediction.

C. BS Deployment Optimization Results … VII. CONCLUSION

The proposed DT-DRL framework progressively learns spatially dispersed BS deployments, efficiently optimizes coverage and joint coverage-capacity objectives, and uses buffer and local-search mechanisms to improve solution quality. Across urban deployment experiments, it substantially reduces computation while approaching idealized performance using geographic data-informed network modeling.

1) Training Analysis:: During training, BS placements evolve from clustered high-reward regions toward spatially dispersed deployments that improve coverage and reduce redundant overlap.This progression occurs as the agent learns to account for inter-BS interference and coverage overlap.
1) Training Analysis:: Exploration rewards fluctuate during the first 1,000 episodes and stabilize around episode 2,000, while the best reward reaches an early plateau.The best episode reward rises rapidly within the first 100 episodes before following a different training pattern.
1) Training Analysis:: The buffer average reward continues increasing after the best reward plateaus, preserving structurally diverse and increasingly superior deployments for later local-search exploitation.The buffer converts the optimizer into a multi-start platform by retaining promising solutions across neighborhoods.
2) Coverage Optimization (β = 1):: DT-Greedy reaches 97.6%, 98.7%, and 98.6% of exhaustive RT performance for M = 4, 5, and 6, respectively, in coverage optimization.All DT-based methods outperform the Hexagonal baseline across deployment scales.
2) Coverage Optimization (β = 1):: DT-Greedy completes coverage optimization in roughly 2 minutes versus approximately 40 hours for exhaustive RT, while achieving comparable performance.The speedup comes from replacing ray tracing with millisecond-level DT evaluations.
3) Joint Coverage-Capacity Optimization (β = 0.5):: With inter-cell interference making the objective non-submodular, the DRL pipeline with local search and buffer consistently surpasses DT-Greedy.The result establishes the need for advanced exploration strategies in joint coverage-capacity optimization.
3) Joint Coverage-Capacity Optimization (β = 0.5):: 0.901 is achieved after LS, Wasserstein buffer, and RT screening, improving from 0.769 with DT-DRL by 17.2% cumulatively.The proposed method takes roughly 22 minutes, while exhaustive RT achieves 0.831 and requires approximately 40 hours; the full pipeline reaches 94.8%–98.9% of genie-aided performance.
VII. CONCLUSION: The framework enables intelligent urban macro BS deployment using only openly available geographic data through a DT combining sample-free radio-map prediction with hybrid geographic and electromagnetic inputs.The hybrid representation captures 3D building heights and material EM properties across kilometer-scale urban areas.