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

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.

💡 Research Summary

The paper tackles the challenging problem of jointly clustering access points (APs), scheduling communication users (CUs) and sensing targets, and selecting the operating mode of each AP in a distributed cell‑free integrated sensing and communication (ISAC) network with limited fronthaul capacity. First, the authors formulate a mixed‑integer linear program (MILP) that captures all relevant constraints: each AP can operate either in transmit or receive half‑duplex mode, the number of served entities cannot exceed the available RF chains, interference among users is mitigated by penalizing highly correlated user pairs, and each scheduled target must be illuminated by at least one transmitting AP and observed by at least one receiving AP. The objective is a weighted sum of normalized communication utility (based on average channel gains) and sensing utility (based on bistatic link gains), with an additional term that rewards preferred AP mode selections. All binary products are linearized with auxiliary variables, yielding a tractable MILP that can be solved by commercial solvers for small‑scale instances.

Because the MILP’s size grows combinatorially with the numbers of APs, users, and targets, solving it in real time is infeasible. To overcome this, the authors propose ASSENT (ASSociation and ENTity selection), a graph neural network (GNN) framework that learns to predict the MILP’s binary decisions from lightweight link statistics only. The network treats APs and entities as nodes in a bipartite graph, attaches features such as average channel gains, spatial correlation matrices, target RCS, distances, and LoS probabilities, and applies message‑passing layers (NNConv, GATv2, TransformerConv were evaluated). The final multilabel classifier outputs the binary variables for AP‑CU association, AP‑target illumination/echo association, AP mode, and target scheduling. Training is supervised using optimal MILP solutions generated offline; class imbalance is handled with weighted cross‑entropy loss.

Extensive simulations with 20–40 APs, 30–60 CUs, and 10–20 targets under various fronthaul constraints demonstrate that ASSENT reproduces the MILP’s trade‑off curve with >95 % fidelity. Communication utility reaches 97 % of the optimal, sensing utility 94 %, and target scheduling success rates are virtually unchanged. Inference time drops from seconds (or minutes) per slot for the MILP to a few milliseconds per forward pass, enabling real‑time operation. Among the GNN variants, GATv2 yields the best accuracy‑complexity balance. The authors also release a full Python/PyTorch implementation, the MILP model, and a dataset of 10 000 random network instances with optimal labels, facilitating reproducibility and future extensions.

The work’s main contributions are: (1) a comprehensive MILP model that jointly handles AP clustering, user/target scheduling, and mode selection under realistic hardware and fronthaul limits; (2) the ASSENT learning‑based framework that achieves near‑optimal performance with orders‑of‑magnitude lower latency; (3) an open‑source codebase and dataset for the community. Limitations include the assumption of static channel statistics and the need for retraining when network conditions deviate significantly from the training distribution. Future research directions suggested are online adaptation, meta‑learning for rapid transfer to new environments, and extension to dynamic mobility scenarios or multi‑service ISAC settings. Overall, the paper presents a solid blend of rigorous optimization and modern machine‑learning techniques to make scalable, low‑latency cell‑free ISAC feasible.