Map2Schedule: An End-to-End Link Scheduling Method for Urban V2V Communications

Lihao Zhang, Haijian Sun, Jin Sun, Ramviyas Parasuraman, Yinghui Ye, Rose Qingyang Hu

TL;DR

Map2Schedule is an end-to-end deep learning framework that generates optimal Vehicle-to-Vehicle (V2V) link scheduling policies using only city maps and vehicle locations. By replacing slow physical channel measurements with a two-stage neural network pipeline, it achieves near-optimal network capacity with a 100x speedup over traditional iterative algorithms.

What Problem Does This Paper Address?

In dense urban environments, V2V communication is a chaotic game of interference. To maximize the "sum-rate" (the total data capacity of all active links), the network must decide which vehicle pairs should transmit and which should stay silent at any given moment.

This is traditionally a "combinatorial integer programming" problem—essentially a complex puzzle that gets exponentially harder as more cars are added. To solve it, you usually need perfect Channel State Information (CSI), which tells you exactly how signals bounce off every building and car. In a high-speed city environment, measuring CSI for every possible interfering link is too slow and overhead-heavy to be practical. Existing fast heuristics often rely only on distance, which fails miserably in cities where a single building can completely block a signal.

Methods

The authors propose a two-stage deep learning approach that bypasses the need for real-time CSI measurements:

  1. CSI Estimation (RadioUNet): The first stage uses a modified RadioUNet, a Convolutional Neural Network (CNN) architecture. It takes a 2D city map and vehicle coordinates as input and predicts "radiomaps"—visual representations of signal strength. The authors specifically addressed common CNN artifacts like "black edges" (mispredicted zero gain near buildings) and "transmitter fading" at map edges to ensure the predicted CSI was physically plausible.
  2. Link Scheduling (Graph Embedding): The predicted CSI is fed into a graph neural network based on structure2vec. Interestingly, they treat each V2V link as a node in the graph. Using mean-field inference, the model propagates interference information across the network to decide the binary state (on or off) for each link.
  3. Weighted Training: Because "good" links (those worth activating) are rare in dense interference scenarios, the authors used a Weighted Binary Cross Entropy (BCE) loss function. This forces the model to prioritize accuracy on high-performing links rather than getting distracted by the "noise" of many low-quality links.

Datasets

  • RadioMapSeer: A massive dataset containing 56,000 radiomaps simulated across 700 different 2D city layouts using high-fidelity ray-tracing and dominant path models (DPM).
  • FPLinQ Baseline: To create "ground truth" labels for training the scheduler, the authors used the FPLinQ algorithm, a high-performing but computationally expensive iterative method.

Key Insights and Ideas

  • Context is King: The paper proves that street layout and vehicle geometry are excellent proxies for physical channel measurements. If the model "knows" there is a building between two cars, it doesn't need to measure the signal to know the interference will be low.
  • Link-Centric Graphing: By representing the communication link itself as a node (rather than the vehicle), the graph embedding naturally captures the relationship between a link's own strength and the interference it receives from neighbors.
  • Scalability: Because the graph embedding model uses shared weights for the classifier, it can handle an arbitrary number of V2V links without needing to be redesigned or retrained for different traffic densities.

Key Results

  • Near-Optimal Performance: Map2Schedule achieved over 90% of the sum-rate performance of the "gold standard" physical-model-based baseline (DPM-FP).
  • Massive Speedup: The system can schedule 50 V2V links in just 0.2 seconds. This is a reduction of two orders of magnitude (100x) compared to traditional iterative baselines, making it viable for real-time use in moving traffic.
  • Superiority over Distance-based Models: While simple distance-based scheduling only achieves about 50% of the potential capacity in urban settings, Map2Schedule’s map-aware approach bridges the gap to near-optimality.
  • Transferability: The model showed "promising transferability," meaning a model trained on short-distance links could adapt to longer-distance scenarios with only a few additional training samples (few-shot learning).

Limitations and Future Work

  • Binary States: Currently, the model only decides if a link is "on" or "off." Real-world systems often use continuous power control, which adds another layer of complexity.
  • Map Dependency: The system relies on having accurate, up-to-date 2D city maps. Performance might degrade if the map data is outdated or if temporary obstacles (like large trucks or construction) aren't accounted for.
  • Transfer Learning Gap: While transfer learning worked, there was still a performance drop when moving between very different scenarios. The authors note that further research is needed to make the model truly "plug-and-play" across diverse global cityscapes.

Summary updated

Abstract

Urban vehicle-to-vehicle (V2V) link scheduling with shared spectrum is a challenging problem. Its main goal is to find the scheduling policy that can maximize system performance (usually the sum capacity of each link or their energy efficiency). Given that each link can experience interference from all other active links, the scheduling becomes a combinatorial integer programming problem and generally does not scale well with the number of V2V pairs. Moreover, link scheduling requires accurate channel state information (CSI), which is very difficult to estimate with good accuracy under high vehicle mobility. In this paper, we propose an end-to-end urban V2V link scheduling method called Map2Schedule, which can directly generate V2V scheduling policy from the city map and vehicle locations. Map2Schedule delivers comparable performance to the physical-model-based methods in urban settings while maintaining low computation complexity. This enhanced performance is achieved by machine learning (ML) technologies. Specifically, we first deploy the convolutional neural network (CNN) model to estimate the CSI from street layout and vehicle locations and then apply the graph embedding model for optimal scheduling policy. The results show that the proposed method can achieve high accuracy with much lower overhead and latency.
BibTeX
@inproceedings{zhang2024map,
  author = {Lihao Zhang and Haijian Sun and Jin Sun and Ramviyas Parasuraman and Yinghui Ye and Rose Qingyang Hu},
  title = {{Map2Schedule: An End-to-End Link Scheduling Method for Urban V2V Communications}},
  booktitle = {ICC},
  year = {2024},
  url = {https://openreview.net/forum?id=792chLhltx},
}

Loading related content...

Figure