PaperlayerView exact revision on arXiv

Source-linked AI summary

Channel-Token Attention for Reliable Dynamic Spectrum Access under Bursty Primary-User Traffic

Krishna Acharya, Dinanath Padhya, Utsab Dahal, Ashish Kandel, Binod Sapkota

arXiv:2608.22992v1cs.NIcs.LGeess.SP

TL;DR

Bursty, class-dependent primary-user activity makes reliable, low-delay dynamic spectrum access difficult. TACAN addresses this with channel tokens carrying occupancy history and AMC entropy plus a QoS context token, and its packet-aware replay reports higher conditional access reliability and lower delay than the compared policies without claiming a throughput gain.

  • Problem

    Bursty, heavy-tailed, class-dependent IoT traffic makes selecting reliable spectrum gaps a difficult sequential decision problem.

  • Method

    TACAN uses a Transformer over per-channel occupancy and AMC-entropy tokens plus a QoS context token, warm-started from an occupancy-greedy policy and refined with PPO.

  • Results

    92.53% ± 0.47 packet-present access success versus 89.94% for Greedy and 83.53% for PPO+MLP, with mean delivery delay of 1.123 slots versus 1.208 for Greedy.

  • Takeaways & Limitations

    Packet-aware replay supports improved access reliability and delay, while delivered packets per SU-slot remain arrival-limited and no packet-throughput gain is claimed.

  • Takeaways & Limitations

    The evaluation uses five seeds, synthetic AMC templates, one trained channel/user configuration, ideal sensing, and entirely simulated controller operation.

Abstract

from arXiv · show

Dynamic spectrum access must coordinate secondary users under bursty primary-user activity while preserving packet reliability and delay. We present TACAN, a centralized policy that represents each channel as a token containing occupancy history and automatic-modulation-classification entropy; a context token supplies queue class, delay and user identity. A Transformer encoder is warm-started from an occupancy-greedy policy and refined with proximal policy optimization. The frozen policies were trained to maintain a channel assignment in every slot, including when queues were empty. We therefore replay them on held-out trajectories and distinguish standby assignment success from packet-present access and packet delivery. In a 20-channel network with 60 primary devices and 4 secondary users, TACAN achieves 92.53% +/- 0.47 packet-present access success, compared with 89.94% for Greedy and 83.53% for PPO+MLP. Its paired gain over Greedy is 2.59 points (parametric 95% CI 1.89-3.29), with wins in all five seeds; the exact two-sided sign-test value is 0.0625. The gain rises from 0.57 points at normal primary-user load to 7.67 points at extreme load. TACAN also reduces mean delivery delay from 1.208 to 1.123 slots and the conditional user-reliability gap from 9.69 to 3.15 points. Delivered packets per SU-slot remain arrival-limited (30.12% versus 30.11%), so no packet-throughput gain is claimed.

I. INTRODUCTION

TACAN targets bursty, class-dependent spectrum access by combining relational channel representations with AMC-derived uncertainty and QoS conditioning. Its evaluation emphasizes packet-aware reliability, delay, load dependence, and paired uncertainty rather than standby assignments alone.

  • Spectrum scarcity and bursty, heavy-tailed, class-dependent IoT traffic make channel selection a difficult sequential decision problem.
  • TACAN represents each channel with occupancy history and AMC entropy, while a QoS context token encodes service class and normalized queueing delay.
  • The Transformer policy is warm-started from an occupancy-greedy heuristic and refined with PPO using class-dependent access and delay rewards, collision penalties, and an URLLC deadline penalty.
  • The controller interface includes a simulation-mode Python decision-path timing measurement, not a live SDN deployment result.
  • Packet-aware replay separates standby assignment success from queued-packet access, delivery, delay, and per-user reliability using frozen policies on identical held-out trajectories.

II. RELATED WORK

Prior work establishes deep reinforcement learning, attention, QoS-aware access, SDN control, and bursty traffic modeling for cognitive radio. TACAN positions its novelty in combining per-channel occupancy and AMC uncertainty with a QoS query token in one shared policy.

  • DRL and attention: Deep reinforcement learning, graph encoders, and attention-based policies address sequential spectrum access and relational structure.
  • Attention in cognitive radio: Prior cognitive-radio attention methods target sensing, users, optical access, power control, or generic resource sequences rather than TACAN’s channel-token representation.
  • PHY-aware access: TACAN feeds AMC-derived uncertainty into channel-selection reinforcement learning, whereas prior AMC work primarily treats classifier output as a report.
  • QoS-differentiated access: QoS-aware spectrum methods use estimated traffic, separate policies, per-class weights, or fixed priority rules; TACAN uses one shared class-conditional policy.
  • SDN and traffic modeling: The work retains SDN control and uses an ON/OFF semi-Markov generator with Bernoulli activation and clipped-Pareto busy periods for bursty traffic.

B. Traffic and PHY Model

The traffic model uses class-specific, heavy-tailed primary-user activity and class-conditioned modulation posteriors. TACAN compresses each channel’s modulation posterior into an entropy feature while acknowledging that its realism remains limited by synthetic templates.

  • Primary-user traffic: Each primary user follows a discrete-time ON/OFF semi-Markov process with probabilistic activation and rounded, clipped Pareto busy durations.
  • Primary-user traffic: The resulting process produces persistent, heavy-tailed busy periods with endpoint clipping, with class parameters reported in Table II.
  • PHY model: Each simulated class has a characteristic distribution over 10 modulation schemes, and the simulator samples class-conditioned AMC posteriors per channel.
  • PHY model: Normalized entropy compresses each 10-dimensional posterior into one permutation-invariant feature that preserves concentration rather than predicted modulation identity.
  • PHY model: The entropy feature is class-correlated in simulation, while validation with a real noisy AMC front end remains outside the evidence.

C. Decision Problem

The decision problem is a partially observable, centralized channel-assignment task with occupancy, modulation-entropy, QoS, delay, and user-identity observations. TACAN restructures this flat observation into shared channel tokens plus a QoS token for attention-based comparison.

  • Decision formulation: The task is modeled as a partially observable Markov decision process in which the controller emits one channel action per secondary user per slot.
  • Observation and action: Each secondary user observes eight-slot occupancy and entropy histories, packet class, normalized queueing delay, and a user identifier; the implementation has 328 input dimensions.
  • Packet-aware evaluation: Assignments are maintained even for empty queues, so standby assignment success is reported separately from access success when a packet is queued.
  • Spectrum tokenisation: Each channel token combines occupancy history, modulation-entropy history, and mean occupancy, giving the network direct access to the greedy baseline statistic.
  • Spectrum tokenisation: A separate QoS token encodes class and delay, positional embeddings form a K + 1 token sequence, and shared attention compares channel representations.

B. Encoder and Heads

TACAN uses Transformer blocks to encode channel and QoS-token representations, then derives policy, value, and auxiliary prediction outputs from the shared representation.

  • The encoder uses pre-norm Transformer blocks with multi-head attention and a two-layer GELU feed-forward network of width 4d.
  • The policy applies a shared linear head to per-channel representations and adds a learnable residual to the greedy occupancy prior.The learned scalar αg controls how strongly the heuristic prior is trusted.
  • The value head pools the QoS token with the channel mean before estimating the state value.
  • An auxiliary head predicts next-slot occupancy for each channel, regularizing the encoder toward predictive representations.Its loss is added with weight 0.1.
  • Attention over K + 1 tokens costs O(L(K+1)^2d), whereas attention over the flattened observation costs O((HKM)^2d).For the reported configuration, the latter is four orders of magnitude larger.

C. QoS-Differentiated Reward

TACAN combines class- and delay-sensitive rewards with warm-started PPO training and a controller pipeline that converts policy assignments into rules, while reported latency remains simulation-only.

  • C. QoS-Differentiated Reward: For a class-c packet, the reward combines success and delay weights with primary-user collision and URLLC deadline penalties.The specified coefficients are β = 1.5, δ = 5.0, and dmax = 50.
  • C. QoS-Differentiated Reward: URLLC receives the largest success weight, a delay weight ten times that of mMTC, and an additional penalty after ten slots.
  • C. QoS-Differentiated Reward: The policy is behaviour-cloned for 20,000 steps against occupancy-greedy before PPO with generalised advantage estimation optimises the clipped surrogate.
  • C. QoS-Differentiated Reward: TACAN and PPO+MLP share the PPO schedule, reward, warm start, and seeds, while TACAN additionally uses auxiliary prediction and a learned greedy-prior term.
  • V. CONTROLLER INTEGRATION: The controller collects simulation state, builds the observation, and converts channel assignments into OpenFlow-like Python rule dictionaries.The default polling interval is 500 ms.
  • V. CONTROLLER INTEGRATION: The reported latency measures only the simulation-mode Python decision path, not live SDN deployment or OpenFlow installation.

VI. EVALUATION

The evaluation replays frozen policies on matched held-out trajectories and separates standby assignment from packet service, showing higher packet-present access success for TACAN without higher arrival-limited throughput.

  • The evaluation uses 20 channels, 60 primary devices, 4 secondary users, five training seeds, and 10,000-step replays on matched held-out seeds.No retraining is performed for the packet-aware analysis.
  • Assignment success includes collision-free standby recommendations for empty queues, whereas packet-present access success counts queued transmission attempts that deliver a packet.
  • TACAN’s legacy assignment-success rate is 92.93%, but 67.59% of its successful assignments occur while the queue is empty.This metric is therefore not packet throughput.
  • TACAN achieves 92.53% ± 0.47 packet-present access success, compared with 89.94% for Greedy and 83.53% for PPO+MLP.
  • Delivered packets per SU-slot are 30.12% for TACAN and 30.11% for Greedy because the fixed 30% arrival process limits offered traffic.The reported claim is improved access reliability and delay, not higher packet throughput.

C. Paired Uncertainty

TACAN’s packet-present access advantage over Greedy is consistent across seeds and grows under extreme primary-user load, while class-level reliability and fairness also improve. The evaluation distinguishes these gains from throughput, which remains arrival-limited.

  • Paired uncertainty: 2.59 points is TACAN’s paired packet-present access gain over Greedy, with a 1.89–3.29-point parametric 95% interval and wins in all 5 seeds.The exact two-sided sign-test value is 0.0625, so the paper emphasizes effect size, interval, and consistency rather than a decisive small-sample significance claim.
  • Behaviour under load: 7.67 points is TACAN’s packet-present access advantage at extreme primary-user load, compared with 0.57 points at normal load.The pattern holds in every seed, although it does not establish a causal mechanism because the complete configurations differ in several ways.
  • Per-class service: 92.56%, 92.53% and 92.50% are TACAN’s packet-present access success rates for URLLC, mMTC and eMBB, respectively.The corresponding Greedy values are 89.72%, 90.01% and 89.99%, and TACAN’s mean delivery delay is lower in all three classes.
  • Fairness: 9.69 to 3.15 points is the reduction in the max–min conditional-access gap, while delivery-rate gaps remain nearly identical at 0.97 versus 0.98 points.The supported conclusion is improved reliability and delay balance, not improved delivered-throughput fairness.

F. What the Attention Learns

The learned QoS-token attention is non-uniform and U-shaped over occupancy rank, emphasizing both least-occupied transmission candidates and most-occupied interferers. Ablations show that input suppression effects depend partly on distribution shift, while AMC information contributes a smaller measured assignment benefit.

  • What the attention learns: U-shaped attention concentrates on the least- and most-occupied channels, while mid-rank channels receive at or below uniform 1/K.The peaks reach roughly twice uniform, increasing from 0.0999 at nominal load to 0.1417 under extreme load.
  • Interpretation: Attention to busy channels may summarize persistence and congestion, but the observed weights alone do not establish a causal mechanism.The paper treats the attention pattern as descriptive.
  • Ablation scope: The 6.82-point assignment-success gap compares complete TACAN and PPO+MLP configurations rather than isolating attention.The models differ in architecture, parameter count, auxiliary loss, and greedy-prior term.
  • Input perturbations: 2.89 points and 4.77 points are the standby assignment-success costs of zeroing modulation entropy and collapsing history to its most recent frame.Zeroing creates an input value never observed during training, so its drop mixes information loss with distribution shift.
  • AMC sensitivity: 0.35 points is the standby assignment-success cost of fully blending a uniform posterior into the simulated AMC output.Because this preserves the feature’s natural range while removing emitter information, it is roughly an order of magnitude smaller than the 2.89-point zeroing cost.

H. Simulation Decision-Path Cost

The simulation-mode controller decision path is measured on one CPU core, with policy inference dominating the reported computation. The benchmark excludes several networking and radio operations, so it is not an end-to-end latency measurement.

  • Measured cost: 5.01 ms median and 17.27 ms at the 95th percentile are the measured decision-path times on one CPU core.The path includes state extraction, policy inference, and rule-dictionary construction.
  • Dominant component: 4.943 ms is the median policy-inference time, making inference the dominant component of the measured decision path.The benchmark reports simulation-mode computation rather than complete networking latency.
  • Scope boundary: The benchmark excludes sensing, AMC inference, OpenFlow serialization and transport, switch rule installation, and radio retuning.Accordingly, the reported figures are not end-to-end networking latency.
  • Interpretation: Packet-aware replay identifies conditional reliability and delay as the discriminating outcomes, while delivered packets per SU-slot remain nearly equal under the 30% arrival process.This separates the packet-present result from a claim of increased offered traffic or packet throughput.

B. What Each Component Contributes

TACAN’s results support the complete configuration as a reliability-oriented policy, while component ablations and evaluation choices limit causal attribution and deployment claims.

  • 9.00 points: TACAN improves packet-present access over PPO+MLP, while mean delay falls by 0.24 slots.The comparison changes tokenisation, attention, auxiliary supervision, the greedy-prior term and capacity.
  • 2.89 standby assignment points: zeroing AMC entropy costs this amount, but destroying its information content costs only 0.35 points.The larger loss is mostly distribution shift rather than lost information; the feature’s value with a real AMC front end is untested.
  • TACAN improves packet-present reliability over Greedy in URLLC, mMTC and eMBB, although the reward, deadline penalty and QoS-token contributions are not isolated.
  • The warm start and learned greedy-prior term were not individually evaluated, so their effects on convergence stability or final reliability are not attributed.
  • 5.01 ms: the measured simulation decision path excludes the live networking path, and the evaluation is not a live SDN deployment result.The model has 485,892 parameters and requires no GPU at inference in the reported setup.
  • Five limitations: post hoc packet-aware replay, five seeds, unmatched PPO+MLP capacity or supervision, synthetic sensing and AMC, one trained configuration, and entirely simulated evaluation constrain the claims.These limitations preclude claims of attention causality, packet-throughput gain, general scaling or deployed SDN performance.
  • 92.53% versus 89.94%: TACAN’s packet-present access success exceeds Greedy, while packet deliveries remain arrival-limited and no throughput gain is claimed.The result supports the complete TACAN configuration within the evaluated simulator.
Loading 2608.22992v1…