Multi-Cell Edge-Intelligent Systems

• Multi-cell edge-intelligent systems are integrated infrastructures combining distributed radio access points, edge servers, and IRS for low-latency AI services.
• They employ joint communication-compute modeling, dynamic service placement, and resource orchestration using optimization and reinforcement learning techniques to boost performance and cost efficiency.
• They enable real-world applications such as autonomous mobility, distributed video analytics, and federated edge learning with significant throughput and latency improvements.

Multi-cell edge-intelligent systems are integrated wireless–compute infrastructures that combine multiple geographically distributed radio access points, edge servers, and often intelligent reflecting surfaces (IRS), to deliver low-latency, high-reliability, and cost-efficient AI-powered services across large regions. Such systems feature tight orchestration of wireless communication, on-site computation, service placement, and resource slicing, leveraging advanced methodologies in stochastic modeling, optimization, and reinforcement learning. The multi-cell architecture enables dynamic multi-agent cooperation, cross-cell resource pooling, and distributed intelligence for applications including autonomous mobility, distributed video analytics, federated learning, and prioritized multi-service orchestration.

1. Network Architecture and Physical Layer Fundamentals

Multi-cell edge-intelligent networks consist of radio cells formed by access points (APs) or base stations (BSs)—each typically co-located with an edge server, and optionally augmented by IRS to assist channel propagation and mitigate blockage issues. Network topology is mathematically modeled using spatial point processes; for example, a homogeneous Poisson point process (PPP) Φ_b of intensity λ_b (BSs/km²) for base station locations and a separate PPP Φ_u for users (Peris et al., 23 Jan 2026, Peris et al., 2022). Each user associates to the nearest BS (minimum path-loss), forming a set of overlapping cells.

Wireless channels experience large-scale path-loss (with exponent α), small-scale Rayleigh fading, fractional power control, and inter-cell interference determined by frequency reuse factor δ. For multi-antenna APs/BSs, maximum ratio combining (MRC) or multi-user detection (MUD) is applied; SINR is expressed as:

$$\mathrm{SINR}_k = \frac{|g_k|^2 \ell(r) r^{-\alpha}}{\sigma^2 + \sum_{z} |g_z|^2 \ell(r_z) d_z^{-\alpha}}$$

where $\ell(r) = \min(P r^{\alpha \epsilon}, \bar{P})$ represents fractional power control (Peris et al., 2022, Peris et al., 23 Jan 2026). IRSs, modeled as programmable passive matrices Φ, manipulate the radio environment by adjusting phase shifts θ_m per element, enhancing effective rates and coverage especially for cell-edge users (Xu et al., 2023, Li et al., 2022).

2. Joint Communication-Compute Modeling and Resource Dimensioning

Multi-cell edge intelligence requires holistic modeling of both wireless and computation resources, as system performance and cost are simultaneously determined by spectrum allocation, AP/BS density, edge CPU/GPU capacity, and queueing under stochastic arrival patterns. For AI inference workloads, frame generation follows a spatial–temporal Poisson process; edge servers implement M/D/1 queues with deterministic inference times $T_s = \frac{c_1 s^3 + c_2}{H}$, where $s$ denotes frame size and $H$ server TFLOPS (Peris et al., 23 Jan 2026, Peris et al., 2022).

The end-to-end offloading delay is given by:

$D = T_{\rm ul} + T_w + T_s$

$T_{\rm ul} = \frac{\theta s^2}{\xi \phi(B, r)}$

where $\phi(B, r)$ is the ergodic rate dependent on system bandwidth $B$, path-loss, number of antennas, and interference regime.

Resource dimensioning is framed as a joint cost-minimization optimization subject to strict tail-latency and accuracy constraints, stability, and coverage, yielding globally optimal deployment configurations via convex reformulation and epigraph splits (Peris et al., 23 Jan 2026). In regimes with different interference/noise limits, densification trades off transmission delay against compute multiplexing.

3. Service Placement, Request Routing, and Resource Orchestration

Multi-cell edge-intelligent systems face the joint challenge of determining service placement (what to cache/run where) and request routing (how user tasks flow to edge servers). The optimization is multi-dimensional:

• Storage: e.g., DNN weights, model caches
• Compute: GHz or CPU cycles/sec per edge node
• Bandwidth: asymmetric uplink/downlink, handled per-request

The problem is formalized as a mixed-binary integer program minimizing cloud offloading subject to per-cell constraints (Poularakis et al., 2019):

$\text{min}~Z = \sum_{u \in \mathcal{U}} y_{l,u}$

with constraints ensuring exact routing, placement feasibility, and bounded resource utilization. Randomized rounding relaxes binary variables to fractions, solves the LP relaxation, and rounds while controlling the probability of resource violations and sub-optimality.

Table: Comparison of Service Placement Approaches

Approach	Resource Utilized	Edge-Hit Fraction
Randomized Rounding	Storage, Compute, BW	Within 10% of LP optimum
Greedy Caching	Storage	10–25% lower
LP Lower Bound	All (fractional)	Optimal

This dimensional optimization enables multi-service edge inference, coded caching, and dynamic capacity adaptation (Poularakis et al., 2019).

4. Multi-Agent Learning, Distributed Control, and SLA Channels

Dynamic resource allocation across multi-cell edge-intelligent systems is enabled by distributed reinforcement learning and multi-agent cooperation. Agents (e.g., per AP, edge server, or radio cell) operate over partitioned action-spaces (cell formation, power allocation, resource-cell definition) (Pervej et al., 2020, Ren et al., 2022).

In the vehicular scenario, a D-MARL approach divides the action-space among agents controlling user–AP associations and transmit powers, with coordination via a central Q-vector. The reward structure enforces energy-efficiency and SINR guarantees, converging rapidly to near-optimal performance and Jain's fairness index ≈0.999 (Pervej et al., 2020).

EdgeMatrix introduces a networked multi-agent actor-critic (NMAC) over resource-cells (bundles of CPU, memory, and latency), clustered into logical SLA channels. Service orchestration and request dispatch (JSORD) are performed at two time-scales, using greedy approximations and linear programming, delivering substantial throughput and SLA compliance improvements—up to +36.7% throughput and 73.7% high-priority SLA edge hits (Ren et al., 2022).

5. IRS-Aided Edge Intelligence and Min–Max Latency Optimization

Intelligent reflecting surfaces are potent enhancements in multi-cell systems, particularly for MEC offloading at cell edges and NLOS regions. The optimization targets joint minimization of energy and latency cost, or worst-case latency, by algorithmically scheduling offload partitions, edge CPU assignments, IRS phase shifts, and user beamforming vectors (Xu et al., 2023, Li et al., 2022).

Block coordinate descent (BCD) alternates between optimization of computation (via QCP), communication (weighted-sum-rate MM/FP/DC), IRS phase design (SDR/SCA), and MUD combining (SOCP), with convergence proofs and fast run-times (<40 outer iterations, <10 mm steps for IRS). IRS deployment at critical regions enables up to 60% reduction in worst-case offloading latency as compared to conventional MEC, with practical recommendations regarding IRS density, CPU split, and trade-offs between computation and communication upgrades (Xu et al., 2023, Li et al., 2022).

Table: IRS-Aided Performance Gains

Method	Latency Reduction	Convergence Speed
IRS (+ BCD-MM)	30–40%	<40 iterations
IRS (+ SDR/SCA)	up to 60% (min–max)	<7 iterations
No IRS	—	—

6. Stochastic Modeling, Video Analytics, and Fairness

For edge video analytics workloads, comprehensive stochastic geometry and queueing analysis reveals two operating regimes: bandwidth-limited (where transmission time dominates) and compute-limited (where server queueing dominates) (Peris et al., 2022, Peris et al., 23 Jan 2026). Coverage probability, ergodic rate, success probability within delay $P_{\text{succ}}$, and effective frame-rate are analytically derived, enabling joint resource adaptation, fairness promotion, and robust performance dimensionaling.

System fairness is quantified via Lorenz curves on $P_{\text{succ}}$ per user distance, with recommendations for central bandwidth allocation to cell-edge users and adaptive frame-size based on instantaneous SINR and queue state.

Resource slicing, admission control, and adaptive scheduling are supported by the closed-form expressions, ensuring reliable, low-latency, and accurate video analytics across large multi-cell deployments.

7. Federated Edge Learning and Over-the-Air Computation in Multi-Cell Scenarios

Federated edge learning (FEEL) benefits from multi-cell wireless aggregation using over-the-air computation (OAC) protocols in both uplink and downlink, specifically employing non-coherent FSK-based majority vote. This approach eliminates channel-state information requirements and stringent synchronization, leveraging inter-cell interference to expand the voting pool and accelerate convergence (Adeli et al., 2022).

Uplink and downlink aggregation occurs synchronously across cells, with signal superposition detected non-coherently, yielding near-uniform accuracy even under heterogeneous data distributions. Convergence to optimality under non-convex loss functions is established; multi-cell OAC outperforms both single-cell OAC and local-only training in speed, fairness, and test accuracy (Adeli et al., 2022).

• Eco-Vehicular Edge Networks for Connected Transportation: A Distributed Multi-Agent Reinforcement Learning Approach (Pervej et al., 2020)
• Energy-Latency Aware Intelligent Reflecting Surface Aided Multi-cell Mobile Edge Computing (Xu et al., 2023)
• Stochastic Modeling and Resource Dimensioning of Multi-Cellular Edge Intelligent Systems (Peris et al., 23 Jan 2026)
• Joint Service Placement and Request Routing in Multi-cell Mobile Edge Computing Networks (Poularakis et al., 2019)
• EdgeMatrix: A Resources Redefined Edge-Cloud System for Prioritized Services (Ren et al., 2022)
• Modelling multi-cell edge video analytics (Peris et al., 2022)
• Multi-cell Non-coherent Over-the-Air Computation for Federated Edge Learning (Adeli et al., 2022)
• Min-Max Latency Optimization for IRS-aided Cell-Free Mobile Edge Computing Systems (Li et al., 2022)