Robust Generalization of Graph Neural Networks for Carrier Scheduling

TL;DR

RobustGANTT tackles the NP-hard problem of carrier scheduling in backscatter IoT networks by deploying a GNN-based Transformer that iteratively assigns node roles to generate complete schedules. The method achieves strong, retraining-free generalization up to 1000 nodes, delivering up to 2x energy/spectrum savings over the best heuristic while maintaining comparable latency, with practical runtimes in the hundreds of milliseconds. Key design insights include the beneficial impact of learning-rate warmup, the superiority of node-degree positional encoding over global encodings, and the optimal use of 12 attention heads. Evaluations on large synthetic topologies and a real 23-node testbed demonstrate consistent resource savings and faster reaction to network changes, underscoring RobustGANTT’s potential for scalable, energy-efficient IoT deployments with backscatter sensor tags.

Abstract

Battery-free sensor tags are devices that leverage backscatter techniques to communicate with standard IoT devices, thereby augmenting a network's sensing capabilities in a scalable way. For communicating, a sensor tag relies on an unmodulated carrier provided by a neighboring IoT device, with a schedule coordinating this provisioning across the network. Carrier scheduling--computing schedules to interrogate all sensor tags while minimizing energy, spectrum utilization, and latency--is an NP-Hard optimization problem. Recent work introduces learning-based schedulers that achieve resource savings over a carefully-crafted heuristic, generalizing to networks of up to 60 nodes. However, we find that their advantage diminishes in networks with hundreds of nodes, and degrades further in larger setups. This paper introduces RobustGANTT, a GNN-based scheduler that improves generalization (without re-training) to networks up to 1000 nodes (100x training topology sizes). RobustGANTT not only achieves better and more consistent generalization, but also computes schedules requiring up to 2x less resources than existing systems. Our scheduler exhibits average runtimes of hundreds of milliseconds, allowing it to react fast to changing network conditions. Our work not only improves resource utilization in large-scale backscatter networks, but also offers valuable insights in learning-based scheduling.

Figures (12)

• Figure 1: RobustGANTT generates schedules for backscatter networks using a GNN-based Transformer model. Step 1: collect MAC and routing protocol information. Step 2: build the IoT network's graph representation, only including edges strong enough for carrier provisioning (e.g., -75 dBm). Step 3: generate the schedule through iterative one-shot node classification. Step 4: disseminate the schedule using existing network flooding mechanisms and append it to the IoT device's normal schedule.
• Figure 2: RobustGANTT has better and more consistent generalization to larger topologies (higher is better). We train eight identical scheduler models for both RobustGANTT (stars) and DeepGANTT perezramirez_DeepGANTT_2023 (squares), and compare them against the TagAlong heuristic PerezPenichet2020afast on larger, previously unseen topologies (without re-training). Isolated markers depict the best performing model, markers joined by lines represent the average, and vertical lines depict standard error.
• Figure 3: Backscatter communication between a tag $\mathtt{T}$ and its host $v_2$ when assisted by neighboring node $v_1$ during a timeslot $s$.
• Figure 4: Example of schedules for a network topology and efficient carrier re-use for concurrent tag interrogation.
• Figure 5: RobustGANTT's ML model architecture. It receives as input a node-feature matrix $\hat{X}_j$ and produces the corresponding schedule timeslot $s_j$ for every iteration $j$. There is a multi-head self-attention GNN in each of RobustGANTT's layers (orange box). $||$ represents a concatenation operation. Green boxes represent a non-linear transformation by a single-layer fully-connected neural network.