Ilmu Komputer & AI editorial
Channel-Token Attention for Reliable Dynamic Spectrum Access under Bursty Primary-User Traffic
The core problem
Dynamic spectrum access (DSA) requires secondary users (SUs) to exploit idle spectrum without harming primary users (PUs). Real PU activity is bursty rather than stationary, so a DSA policy must track rapidly changing occupancy while simultaneously preserving packet reliability and delay. The authors identify a measurement problem in prior learning-based DSA evaluations: policies are often trained to emit a channel assignment in every slot, including slots where the SU queue is empty. Such standby assignments can inflate apparent success rates and obscure whether a policy actually delivers packets when a packet is present.
This paper introduces TACAN (Token-Attention Channel Allocation Network), a centralized policy that treats each channel as a token. A channel token encodes occupancy history together with automatic-modulation-classification (AMC) entropy, while a separate context token supplies queue class, delay, and user identity. The Transformer encoder is warm-started from an occupancy-greedy policy and then refined with proximal policy optimization (PPO). The stated goal is reliable dynamic spectrum access under bursty PU traffic, evaluated with a protocol that separates st
Innovation
Experiments use 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. The 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 advantage is load-dependent. The gain rises from 0.57 points at normal primary-user load to 7.67 points at extreme load, indicating that channel-token attention matters most precisely when PU traffic is most bursty. TACAN also reduces mean delivery delay from 1.208 to 1.123 slots and shrinks the conditional user-reliability gap from 9.69 to 3.15 points.
Delivered packets per SU-slot remain arrival-limited: 30.12% for TACAN versus 30.11% for the comparison, so the authors explicitly claim no packet-throughput gain. The improvement is in access reliability, delay, and fairness across users, not in raw delivered volume.
Why it matters
The results support a specific interpretation: representing channels as tokens with occupancy history and AMC entropy gives the policy a richer state abstraction than a flat observation vector, and this abstraction pays off under bursty PU activity. The load-dependent gain—0.57 points at normal load rising to 7.67 points at extreme load—is consistent with attention over channel tokens helping the policy disambiguate which channels are transiently available when occupancy statistics shift quickly.
The reduction of the conditional user-reliability gap from 9.69 to 3.15 points is arguably as important as the headline access-success number, because it indicates that the gains are not concentrated in a single favored SU. The delay reduction from 1.208 to 1.123 slots is modest but directionally consistent.
The arrival-limited throughput result (30.12% versus 30.11% delivered packets per SU-slot) is a deliberate negative finding. It constrains the claim: TACAN improves the reliability and timeliness of access under bursty PU traffic, but it does not create additional traffic to deliver. The sign-test value of 0.0625 across five seeds is also a reminder that the evidence, while consistent in direction, is based on a small number of seeds.
Methodologically, the paper's separation of standby assignment success from packet-present access and packet delivery is a reusable contribution. Any learned DSA policy trained to always emit an assignment should be evaluated on held-out trajectories with this decomposition, otherwise empty-queue assignments can inflate reported success.
Who should read this
Opening member content…