Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ASSENT: Learning-Based Association Optimization for Distributed Cell-Free ISAC

Mehdi Zafari, A. Lee Swindlehurst

TL;DR

This work tackles scalable joint association in distributed cell-free ISAC under fronthaul constraints by formulating a MILP that jointly optimizes AP clustering, user/target scheduling, and AP mode with RF-chain limits and sensing requirements. To enable real-time operation, it introduces ASSENT, a graph neural network trained on offline MILP solutions to predict near-optimal associations from lightweight link statistics, enabling a single forward-pass inference. The authors provide an open-source Python/PyTorch implementation and comprehensive datasets, demonstrating that ASSENT achieves near-MILP utility with accurate, scalable decisions and reduced latency. The study shows that learning-based association can closely emulate optimal MILP decisions while enabling practical deployment in large-scale, distributed ISAC networks. Future work includes reinforcement learning for continual adaptation and experimental validation on distributed ISAC testbeds.

Abstract

Integrated Sensing and Communication (ISAC) is a key emerging 6G technology. Despite progress, ISAC still lacks scalable methods for joint AP clustering and user/target scheduling in distributed deployments under fronthaul limits. Moreover, existing ISAC solutions largely rely on centralized processing and full channel state information, limiting scalability. This paper addresses joint access point (AP) clustering, user and target scheduling, and AP mode selection in distributed cell-free ISAC systems operating with constrained fronthaul capacity. We formulate the problem as a mixed-integer linear program (MILP) that jointly captures interference coupling, RF-chain limits, and sensing requirements, providing optimal but computationally demanding solutions. To enable real-time and scalable operation, we propose ASSENT (ASSociation and ENTity selection), a graph neural network (GNN) framework trained on MILP solutions to efficiently learn association and mode-selection policies directly from lightweight link statistics. Simulations show that ASSENT achieves near-optimal utility while accurately learning the underlying associations. Additionally, its single forward pass inference reduces decision latency compared to optimization-based methods. An open-source Python/PyTorch implementation with full datasets is provided to facilitate reproducible and extensible research in cell-free ISAC.

ASSENT: Learning-Based Association Optimization for Distributed Cell-Free ISAC

TL;DR

This work tackles scalable joint association in distributed cell-free ISAC under fronthaul constraints by formulating a MILP that jointly optimizes AP clustering, user/target scheduling, and AP mode with RF-chain limits and sensing requirements. To enable real-time operation, it introduces ASSENT, a graph neural network trained on offline MILP solutions to predict near-optimal associations from lightweight link statistics, enabling a single forward-pass inference. The authors provide an open-source Python/PyTorch implementation and comprehensive datasets, demonstrating that ASSENT achieves near-MILP utility with accurate, scalable decisions and reduced latency. The study shows that learning-based association can closely emulate optimal MILP decisions while enabling practical deployment in large-scale, distributed ISAC networks. Future work includes reinforcement learning for continual adaptation and experimental validation on distributed ISAC testbeds.

Abstract

Integrated Sensing and Communication (ISAC) is a key emerging 6G technology. Despite progress, ISAC still lacks scalable methods for joint AP clustering and user/target scheduling in distributed deployments under fronthaul limits. Moreover, existing ISAC solutions largely rely on centralized processing and full channel state information, limiting scalability. This paper addresses joint access point (AP) clustering, user and target scheduling, and AP mode selection in distributed cell-free ISAC systems operating with constrained fronthaul capacity. We formulate the problem as a mixed-integer linear program (MILP) that jointly captures interference coupling, RF-chain limits, and sensing requirements, providing optimal but computationally demanding solutions. To enable real-time and scalable operation, we propose ASSENT (ASSociation and ENTity selection), a graph neural network (GNN) framework trained on MILP solutions to efficiently learn association and mode-selection policies directly from lightweight link statistics. Simulations show that ASSENT achieves near-optimal utility while accurately learning the underlying associations. Additionally, its single forward pass inference reduces decision latency compared to optimization-based methods. An open-source Python/PyTorch implementation with full datasets is provided to facilitate reproducible and extensible research in cell-free ISAC.

Paper Structure

This paper contains 13 sections, 8 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. System and Channel Model
  3. Communication Channel Model
  4. Sensing Channel Model
  5. Association Optimization Problem
  6. Optimization Variables
  7. Objective Function
  8. MILP Formulation
  9. Learning-Based ASSociation and ENTity Selection: ASSENT
  10. Dataset Generation
  11. GNN Architecture
  12. Evaluation
  13. Conclusion

Figures (2)

  • Figure 1: Performance of the MILP-based optimization \ref{['eq:opt_problem']}. (Left) Empirical CDF of the MILP utility and four greedy baselines: MILP-aligned (schedules exactly the same number of users and targets as MILP but greedily), channel-only (greedy only on channel gain), comm-only (scheduling only CUs greedily), and sens-only (scheduling only targets). (Right) Variation of network metrics: UE coverage fraction, scheduled-target fraction, Tx-AP fraction, and communication RF-share versus the trade-off parameter $\alpha$. Shaded band denotes the inter-quartile range (25th-75th percentile).
  • Figure 2: Performance of ASSENT. (Left) Training (solid) and validation (dashed) loss for three GNN architectures, with learning-rate schedule shown on the right y-axis. (Right) Average objective values on the test set achieved by each GNN architecture compared with MILP benchmark.