Toward Integrated Sensing, Communications, and Edge Intelligence Networks

1. Problem Statement

The paper addresses the emerging challenge of managing resource allocation in sixth-generation (6G) wireless networks where three distinct services must coexist on the same infrastructure:

1. Integrated Sensing and Communications (ISAC): Using radio-frequency signals simultaneously for data transmission and environmental sensing (e.g., radar).
2. Edge Intelligence (EI): Processing data (e.g., video from cameras) at the network edge using Artificial Intelligence (AI) models for tasks like object classification.
3. Goal-Oriented Semantic Communications (SemCom): Transmitting only the information necessary to achieve a specific task goal, rather than raw data fidelity.

The Core Conflict:
Current approaches often treat these services in isolation or assume disjoint resource allocation. However, there is a fundamental cross-layer coupling:

• Computation vs. Communication: The choice of an AI model (e.g., lightweight vs. heavy) dictates the computational delay ($L_{comp}$). This directly impacts the remaining time budget available for data transmission ($L_{comm}$), which in turn affects the required uplink bandwidth and power.
• Communication vs. Sensing: The time and power allocated to the uplink (for EI data) reduce the resources available for the downlink (ISAC). Since ISAC performance (sensing accuracy) depends heavily on downlink power and duration, optimizing for EI degrades sensing, and vice versa.

The paper seeks to solve the joint optimization problem of allocating wireless resources (power, time) and selecting AI model configurations to maximize the "Goal Effectiveness" of the EI service while satisfying ISAC sensing constraints and communication rate requirements.

2. Methodology

System Model

• Architecture: A Full-Duplex (FD) Base Station (BS) with $N_t$ transmit and $N_r$ receive antennas.
• Services:
  • Uplink (UL): A camera uploads compressed image batches to an Edge Server (ES) for inference.
  • Downlink (DL): The BS serves a user (e.g., video streaming) while simultaneously sensing a target using the DL signal echoes.
• Time Structure: Time is divided into frames. A portion $\rho_{ul}$ is dedicated to UL (EI), and the remainder $\rho_{dl}$ is dedicated to ISAC (DL + Sensing).
• Constraint: The BS operates in "quasi-FD" mode. While it can transmit and receive simultaneously for ISAC, it must switch to half-duplex (receive-only) during UL data uploads because the strong camera uplink signal would overwhelm the weak radar echoes.

Key Performance Indicators (KPIs)

1. Inference Delay ($L_{tot}$): Sum of communication delay (upload time) and computation delay (model inference time).
   • $L_{tot} = L_{comm} + L_{comp}$.
   • $L_{comp}$ depends on the selected ML model (e.g., MobileNet vs. ViT).
2. Goal Effectiveness ($E_g$): A probabilistic metric defined as the likelihood that a task is successful within a maximum delay ($L_{max}$) and minimum quality ($Q_{min}$).
   • Success requires: $Q(nb, m) \ge Q_{min}$ AND $L_{tot} \le L_{max}$.
3. ISAC Metrics:
   • Achievable Rate: Downlink throughput for the user.
   • Sensing Accuracy: Measured by the Cramér-Rao Bound (CRB) for target angle estimation. Lower CRB implies higher accuracy.

Optimization Formulation

The authors formulate a Mixed-Integer Non-Linear Programming (MINLP) problem ($P$):

• Objective: Minimize a weighted sum of BS transmit power (for ISAC) and the "richness" of the data representation (number of bits $n_b$).
  • Minimizing power saves energy; maximizing bits improves inference quality.
• Constraints:
  • Sensing: CRB must be below a threshold ($CRB_{th}$).
  • Communication: Downlink rate must meet a threshold ($R_{th}$).
  • Power: Total DL power $\le P_{max}$.
  • Latency: Implicitly handled via the relationship between computation time and available transmission time.
• Solution Approach:
  • The problem is non-convex due to the integer variable $n_b$ (data bits).
  • However, for a fixed $n_b$, the problem becomes convex.
  • The authors solve the problem by iterating through the finite set of possible data representations ($N_b$), solving the convex sub-problem for each using CVXPY, and selecting the global optimum.

3. Key Contributions

1. Triple-Functional Network Concept: The paper introduces a framework where ISAC, Communications, and Edge Intelligence are jointly optimized, explicitly modeling the trade-offs between them.
2. Computation-Aware Resource Allocation: Unlike traditional schemes that treat computation as a separate layer, this work integrates AI model complexity (GFLOPs) directly into the physical layer resource allocation (power and time).
3. Analytical Coupling: The authors derive explicit mathematical expressions showing how the choice of an AI model ($L_{comp}$) constrains the time available for ISAC ($\rho_{dl}$), which subsequently dictates the sensing accuracy (CRB) and communication rate.
4. Optimal Algorithm: A solution method is provided that optimally balances the trade-off between DL transmit power and EI goal effectiveness.

4. Numerical Results and Discussion

The authors simulated a scenario with a 10 GHz carrier, 50 MHz bandwidth, and image classification on CIFAR-100.

• Models Tested: MobileNet-v3-small (0.12 GFLOPs), ResNet-50 (8.18 GFLOPs), and ViT-B-16 (33.7 GFLOPs).
• Benchmark: Compared against a "computation-unaware" strategy that assumes the most complex model (ViT) regardless of the actual resources available.

Key Findings:

• Power-Efficiency Trade-off: The proposed computation-aware approach significantly outperforms the benchmark. For example, using a medium-complexity model (ResNet-50) with the proposed method achieves the same "Goal Effectiveness" as the benchmark using the heavy model (ViT) but with ~50% less DL transmit power.
• Flexibility: The system can adapt to sensing constraints. If the sensing requirement (CRB) is relaxed, the system can allocate more time to the uplink, allowing for higher data resolution and better inference quality without increasing power.
• Model Selection: The results show that for high transmit powers (>80 mW), lighter models (ResNet-50) can achieve >80% goal effectiveness, making the heavy ViT model unnecessary and wasteful of resources.

5. Significance

This work is significant for the development of 6G networks for several reasons:

• Holistic Design: It moves beyond the siloed optimization of "sensing vs. comms" or "comms vs. compute," recognizing that in an AI-driven network, these are inextricably linked.
• Resource Efficiency: By dynamically selecting the appropriate AI model based on available wireless resources, networks can drastically reduce energy consumption and latency while maintaining service quality.
• Practical Implementation: The "quasi-FD" architecture proposed is realistic, acknowledging the physical limitations of separating strong uplink data from weak radar echoes, while still leveraging FD capabilities for ISAC.
• Future Directions: The paper lays the groundwork for dynamic, context-aware networks where AI model selection and resource allocation happen in real-time based on target behavior and channel conditions.

In summary, the paper demonstrates that ignoring the computational cost of AI models when allocating wireless resources leads to suboptimal performance, and that a joint, cross-layer optimization strategy is essential for efficient 6G ISAC-EI networks.