In satellite constellations and vehicle networks, transmitters must choose power levels and link schedules while the pattern of who interferes with whom changes every few milliseconds. The fastest learned controllers for this problem are graph neural networks, meaning networks that pass messages along the edges of a graph and here treat each link as a node and each interference path as an edge. They solve one frozen snapshot of the network at a time. Streaming attention, developed for live speech recognition, is a different tool: a model decides online while it caches summaries of the past and peeks a short, fixed distance into the future. In speech, every frame of peeking adds latency. Moving networks differ in one useful way. The large-scale part of the future channel, which is distance and path loss set by orbits or planned routes, is largely known in advance from orbital data or navigation. Peeking at that part of the future costs nothing.
This proposal builds a streaming attention layer on top of an interference graph network. At each step, every link attends to two sources: cached embeddings of its own past measured channels, and look-ahead tokens computed from the predicted future geometry. The goal is to make power and scheduling decisions that anticipate interference changes and do not thrash when the topology changes. The claim is narrow and testable. When decisions are coupled over time, either because channel measurements arrive late or because reconfiguring links has a cost, a controller that combines cached history with free geometric look-ahead should beat both per-snapshot graph controllers and recurrent graph controllers. The experiments are also designed to show whether the attention mechanism itself matters, or whether the look-ahead information alone does the work.
Core Claim
Snapshot controllers can only be beaten in specific regimes, so those regimes need to be stated first. If current channel state is perfect and there is no cost to changing decisions, then solving each snapshot optimally is optimal overall, and no temporal model can help. The proposal therefore targets the objective
$$\max_{\pi}\ \mathbb{E}\Big[\sum_t \sum_i \log\big(1+\mathrm{SINR}_i(p_t, H_t)\big) - \lambda \sum_t \|p_t - p_{t-1}\|_1 - \mu \cdot \#\mathrm{handovers}_t\Big], \quad p_t = \pi(H_{\le t-d}, \hat G_{t:t+R})$$
The symbols are defined as follows. $p_t$ is the vector of transmit powers or link activations at step $t$. $H_t$ is the true channel matrix, and $\mathrm{SINR}$ is the signal-to-interference-plus-noise ratio. $\lambda$ and $\mu$ price power reconfiguration and handovers. $d$ is the channel-measurement delay. $\hat G_{t:t+R}$ is the predicted large-scale gain matrix over the next $R$ steps, obtained from satellite orbital elements or vehicle routes.
The claim is that for $d>0$ or $\lambda,\mu>0$, a policy using both cached history and $\hat G$ outperforms per-snapshot policies. It should also match or beat a recurrent policy given the same $\hat G$. The attention form should help most when topology changes are frequent: interferers disappear and reappear, and links are born and die. A fixed-size recurrent state has to overwrite old relationships in exactly that situation, whereas an attention cache can retrieve them.
Research Program
The spatial part reuses existing designs. Its message-passing layers follow the wireless channel graph convolution network of Graph Neural Networks for Scalable Radio Resource Management, which shows that such networks can represent distributed optimizers such as the classical weighted minimum mean-squared-error algorithm (WMMSE). They transfer from 50 to 1,000 links with about a 1–2% loss. Training is unsupervised and uses the primal-dual method of Optimal Wireless Resource Allocation With Random Edge Graph Neural Networks, which keeps power budgets satisfied without labels.
The temporal part adapts two speech mechanisms. It takes block processing with cached left-context keys and values plus a compact memory bank from Emformer (Transformer in action). It takes the pattern of unlimited left context, finite right look-ahead, and reuse of state across chunks from MFLA (MFLA). In our layer, each link node keeps a cache of its last $W$ spatial embeddings. At each step it attends to three things: that cache, $R$ look-ahead tokens built by running the same graph layers on $\hat G$, and a memory-bank summary. Per-step cost is therefore $O(W+R)$ per node, with no recomputation of the window. That is the specific inefficiency of the sliding-window temporal convolution in GATE-HPPO. The cache is keyed by neighbor identity, so when an interfering satellite beam returns, the attention can retrieve that pair's earlier history.
Two findings in the file shape and limit the design. ILCP reports that an unconstrained Transformer produced a 22.6% ping-pong handover rate, against 0.0% for its recurrent approach. We read this as a warning, not a refutation. Our attention is strictly causal on measured data, and it sees the future only through deterministic geometry. Whether causality and bounded context remove that failure is one of the checks below. The second limit is that per-snapshot graph controllers are already very fast: 6 ms for 1,000 pairs in the channel graph network paper. Any temporal model must therefore justify its additional cost with gains on the time-coupled objective, not with speed.
The environments are three. The first adds mobility to the Rayleigh-fading ad-hoc setting of Unfolding WMMSE Using Graph Neural Networks. We use path-loss exponent 2.2, train at 20 pairs and test at 10 to 30 or more, and move nodes on known random-waypoint routes, with Gauss-Markov small-scale fading as the unknown component. The second is a co-channel multi-beam low-Earth-orbit downlink built from the Starlink G6-35 orbital setting of Service Function Chain Dynamic Scheduling in SAGIN and the OneWeb orbital-element setting of SkyLink. Here we add the cross-beam interference that SkyLink omits. The third is the Vienna 4G/5G drive-test trace from ILCP, used only for the handover check. It contains 31 test handovers, so we will report confidence intervals and treat the result as indicative.
Baselines
The naive reference is a per-snapshot WMMSE, or a per-snapshot WCGCN, applied to the latest available but delayed channel measurements. This is how current learned controllers would actually be deployed. In the static setting, the relevant published figures are 83.21 average sum-rate for 4-layer UWMMSE versus 82.94 for classical WMMSE at 20 pairs (UWMMSE), and 97–103% of WMMSE's sum-rate for WCGCN (Shen et al.). We did not find a reported result on Lacuna for any learned controller on a mobile interference benchmark with delayed channel measurements or switching costs. We will therefore reproduce UWMMSE and WCGCN in our environments with the same training steps and hyperparameter search budget as our model.
The strongest competitor is a recurrent graph controller. It uses the same WCGCN layers followed by a per-node gated recurrent unit, which mirrors the ILCP design (ILCP). It is strong because it is causal, keeps constant state, costs little, and already beats a Transformer on handovers. A second strong competitor is warm-started WMMSE, which initializes each step from the previous solution and runs a fixed number of iterations matched to our latency.
For all baselines we fix the channel-measurement delay, the number of parameters (within 20%), wall-clock latency on the same GPU, and the tuning budget. Beating a baseline means a statistically significant gain in the time-coupled objective over 6,400 or more test trajectories, with no loss of zero-shot size transfer. For handovers, it means a ping-pong rate at or below the recurrent baseline's together with higher post-handover accuracy.
Two ablations isolate the key ingredients. One removes the look-ahead tokens, giving causal streaming attention only. The other keeps look-ahead but replaces attention with the recurrent unit, which receives $\hat G$ as extra input features.
Why not just…?
Why not just run the fast snapshot GNN more often? Faster updates shrink the measurement delay $d$, but they cannot push it below the measurement and feedback latency. Faster updates also make reconfiguration churn worse, not better. The "do nothing" baseline evaluated at several update rates will show whether any affordable rate closes the gap. If it does, the proposal's claim fails for that regime, and we will report that.
Why not just add a GRU to the graph network? This is the real threat. ILCP shows that recurrence can be very effective, and the file contains no evidence that attention beats it on wireless data. Our case rests on topology churn, where a fixed state must overwrite the history of interferers that later return. We will sweep churn rate, meaning the fraction of interference edges born or killed per step, and state in advance the expected outcome: roughly equal performance at low churn and an attention advantage at high churn. If the recurrent unit given look-ahead features matches us at every churn level, we will conclude that the useful ingredient is free geometric look-ahead, not attention.
Why not just predict the next channel and hand it to a snapshot solver? A predict-then-optimize pipeline can use $\hat G$ too. However, it optimizes each step in isolation, so it still pays switching and handover costs that it never anticipates. We run this baseline explicitly: a learned channel predictor followed by WCGCN, given the same $\hat G$. If the time-coupled terms $\lambda$ and $\mu$ are zero and this baseline ties us, that confirms the regime analysis in the core claim, not a failure of the method.
Proposed Experiments
The first decisive check is the regime map. We sweep measurement delay $d$ from 0 to 20 steps and switching prices $\lambda,\mu$ from 0 upward. We expect no gain at $d=0$ with zero switching prices, and a growing gain as either increases. A gain at $d=0$ with zero prices would indicate a bug or leakage.
The second check degrades the look-ahead by adding orbital or route prediction error to $\hat G$. This tests whether the benefit decays gracefully back to the causal-only ablation, or whether bad forecasts make the model worse than having no look-ahead.
The third check is the churn sweep against the recurrent baseline described above.
The fourth check is the handover test on the Vienna trace, against the published figures of 0.0% ping-pong for ILCP, 6.5% with no context, and 22.6% for the Transformer. It asks whether bounded causal attention avoids the Transformer collapse.
Finally, we measure cost per step, in operations and latency, against a sliding-window temporal convolution in the style of GATE-HPPO, and zero-shot transfer from 20 to 30, and from 50 to 500, links or beams.
References
- Graph Neural Networks for Scalable Radio Resource Management: Architecture Design and Theoretical Analysis
- Optimal Wireless Resource Allocation With Random Edge Graph Neural Networks
- Transformer in Action: A Comparative Study of Transformer-Based Acoustic Models for Large Scale Speech Recognition Applications
- MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition
- Graph-Aware Temporal Encoder-Based Service Migration and Resource Allocation in Satellite Networks
- Inductive Latent Context Persistence: Closing the Post-Handover Cold Start in 6G Radio Access Networks
- Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation
- Service Function Chain Dynamic Scheduling in Space-Air-Ground Integrated Networks
- SkyLink: Scalable and Resilient Link Management in LEO Satellite Networks