⚡ electrical engineering

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

The paper introduces TACAN, a Transformer-based dynamic spectrum access policy that leverages channel tokens and context information to significantly outperform greedy and standard reinforcement learning baselines in packet-present access success and reliability under bursty primary-user traffic, particularly at extreme load levels.

Original authors: Krishna Acharya, Dinanath Padhya, Utsab Dahal, Ashish Kandel, Binod Sapkota

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Krishna Acharya, Dinanath Padhya, Utsab Dahal, Ashish Kandel, Binod Sapkota

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The air around us is filled with invisible waves carrying our messages, from the text on a phone to the data streaming to a smart home. For decades, the rules for using these waves have been strict and static, like assigning a specific lane on a highway to a single driver who never leaves. But as the number of connected devices explodes, this rigid system is becoming a bottleneck. The solution lies in a concept called dynamic spectrum access, where secondary devices are allowed to temporarily use a frequency band only when the primary, licensed owner is not using it. The challenge is that these primary users often appear and disappear in unpredictable bursts, much like sudden traffic jams. If a secondary device guesses wrong and transmits while the primary user returns, the signal is lost, and the data is corrupted. Making the right choice in a split second, among dozens of available channels, is a complex puzzle that traditional computers struggle to solve efficiently.

A team of researchers at Tribhuvan University in Nepal has developed a new approach to this puzzle, creating a system they call TACAN. Instead of treating the available radio channels as a simple list of options, the researchers taught a computer program to view them as a sequence of distinct items, each with its own history and personality. In their simulation, the system looks at the past activity of each channel and also analyzes a subtle statistical clue derived from the type of signal currently occupying it. This clue, known as entropy, acts like a measure of uncertainty or confusion in the signal, helping the system predict how long a channel might stay busy. The system then combines this information with the specific needs of the message it is trying to send. A message that must arrive instantly gets treated differently than one that can wait, allowing the system to prioritize urgent data without needing a separate set of rules for every situation.

The core of this innovation is a type of artificial intelligence architecture that allows the system to compare all available channels against one another simultaneously. Imagine a traffic controller who does not just look at one road at a time but can instantly see how a jam on one street affects the flow on all others. By using this method, the system learns to recognize patterns in the chaotic traffic of wireless signals. The researchers trained this system using a method where it first learned from a basic, simple strategy before refining its skills through trial and error. They tested it in a simulated environment with twenty different channels, sixty primary devices creating unpredictable traffic, and four secondary users trying to send data. The results showed that this new method was significantly more reliable than older, simpler strategies. When a packet of data was ready to be sent, the new system successfully found a clear channel 92.53 percent of the time, compared to about 90 percent for the standard greedy approach and 83 percent for other advanced learning methods.

Perhaps more importantly, the system reduced the time it took for data to reach its destination. On average, messages waited only 1.123 time slots before being delivered, a noticeable improvement over the 1.208 slots required by the standard method. The system also proved to be much fairer, ensuring that no single user was left waiting significantly longer than the others. The researchers found that the advantage of their new system grew larger as the network became more crowded. Under normal conditions, the improvement was small, but when the network was under extreme load, the new system outperformed the standard method by nearly eight percentage points. This suggests that the ability to understand the relationships between different channels becomes most valuable when the airwaves are most congested.

The study also carefully examined what parts of the system were actually doing the heavy lifting. They discovered that the statistical clue about signal uncertainty, while helpful, was not the main driver of success. The real power came from the way the system organized its view of the channels and how it used a specific token to represent the needs of the message. The researchers were careful to note that these results come from a computer simulation, not a live radio network. They did not claim that the system would instantly increase the total amount of data that could be sent, because the number of messages arriving was fixed by the simulation. Instead, the achievement is one of reliability and speed: the system ensures that when a message needs to go, it gets there with fewer errors and less delay.

In the end, this work offers a clearer path for managing the crowded airwaves of the future. By treating the spectrum not as a flat list of options but as a dynamic, interconnected landscape, and by giving the system a way to understand the urgency of each message, the researchers have shown a way to make wireless communication more robust. While the technology is still in the testing phase within a simulated world, the results suggest that the next generation of wireless networks could be much more efficient at handling the unpredictable bursts of traffic that define our connected lives. The system does not solve the problem of scarcity, but it makes the best possible use of the space that is available, ensuring that critical data finds its way through the noise.

Technical Summary: TACAN for Dynamic Spectrum Access

Problem Statement
Dynamic Spectrum Access (DSA) in dense IoT environments faces the challenge of coordinating secondary users (SUs) under bursty, heavy-tailed primary user (PU) traffic while maintaining packet reliability and low latency. Traditional Deep Reinforcement Learning (DRL) approaches often flatten channel history into vectors, losing relational structure, or rely on graph encoders that model user interference rather than per-channel temporal and physical-layer (PHY) observations. Furthermore, existing policies frequently fail to distinguish between "standby" channel assignments (when queues are empty) and actual packet transmission, obscuring true access reliability. The core difficulty lies in making sequential decisions on which spectrum gaps to utilize when traffic is class-dependent (e.g., URLLC vs. mMTC) and the environment is partially observable.

Methodology: TACAN Architecture
The paper proposes TACAN (Token-based Attention for Channel Assignment Network), a centralized policy that treats the spectrum as a sequence of tokens rather than a flat state vector.

Tokenization:
- Channel Tokens: Each of the $K$ channels is represented as a token containing its occupancy history over the last $H$ slots and a per-channel Automatic Modulation Classification (AMC) entropy. This entropy is derived from the normalized posterior distribution of modulation classes, serving as a compact uncertainty statistic correlated with PU holding times.
- QoS Context Token: A single additional token encodes the secondary user's current packet class (e.g., URLLC, mMTC, eMBB), normalized queueing delay, and user identity.
Transformer Encoder:
- A Transformer encoder processes the $K+1$ tokens. This allows channel tokens to attend to one another (capturing spectrum-wide load and persistence) and allows the QoS token to attend over channels.
- This structure enables a single shared policy to dynamically re-rank channels based on the specific service class of the packet in flight, without needing separate networks for different traffic types.
- Complexity: The attention cost is $O(L(K+1)^2d)$ , significantly lower than applying attention to a flattened observation of size $HKM$ .
Policy and Training:
- Warm-Start: The policy is initialized via behavior cloning from an occupancy-greedy heuristic to avoid the random-exploration phase of standard PPO.
- Refinement: The policy is refined using Proximal Policy Optimization (PPO).
- Architecture Details: The policy head includes a learnable residual term based on the greedy occupancy prior, allowing the agent to learn how much to trust the heuristic. An auxiliary head predicts next-slot occupancy to regularize the representation.
- Reward: The reward function includes class-dependent access weights, delay penalties, PU collision penalties, and specific deadline penalties for URLLC traffic.
Evaluation Protocol:
- The authors introduce a packet-aware re-evaluation method. They replay frozen policies on held-out trajectories to distinguish between "standby assignment success" (empty queues) and "packet-present access success" (actual transmission attempts). This prevents the inflation of metrics by successful assignments to empty queues.

Key Contributions

Channel Tokens with QoS Query: The paper introduces a representation where each channel token fuses temporal occupancy with AMC-derived entropy, conditioned by a QoS token. This allows a single policy to adapt to different service classes via the attention mechanism.
Packet-Aware Checkpoint Re-evaluation: The study separates continuous standby metrics from actual packet delivery metrics, providing a more accurate assessment of access reliability.
Load-Dependent Reliability Analysis: The evaluation reports paired effect sizes and exact sign tests across five independently trained seeds, analyzing performance under normal, high, and extreme PU loads.

Results
In a simulated 20-channel network with 60 PUs and 4 SUs:

Access Success: TACAN achieves 92.53% ± 0.47 packet-present access success, outperforming the Greedy baseline (89.94%) and a PPO+MLP baseline (83.53%).
Gain Magnitude: The gain over Greedy is 2.59 percentage points (95% CI: 1.89–3.29). This gain is load-dependent, rising from 0.57 points at normal load to 7.67 points at extreme load.
Delay and Fairness: TACAN reduces mean delivery delay from 1.208 to 1.123 slots and significantly narrows the conditional user-reliability gap from 9.69 to 3.15 points.
Throughput: Due to the fixed 30% packet arrival rate, delivered packets per SU-slot remain nearly identical across methods (~30.1%), meaning the improvement is in reliability and delay, not raw throughput.
Ablation Insights:
- Removing the AMC entropy feature (by zeroing it) causes a 2.89-point drop, but analysis suggests most of this is due to distribution shift. Blending a uniform posterior (preserving the range but removing information) costs only 0.35 points, indicating the AMC signal provides a modest but genuine information gain.
- The comparison with PPO+MLP is a full configuration comparison (differing in architecture, capacity, and auxiliary losses), not an isolated attention ablation.

Significance and Claims
The paper claims that TACAN demonstrates the value of tokenizing the spectrum to combine relational channel structure with PHY-aware uncertainty statistics (AMC entropy) and QoS conditioning.

The primary contribution is a reliability-oriented channel assignment policy that adapts to traffic class and load conditions more effectively than flat encoders or greedy heuristics.
The authors explicitly state that they do not claim a gain in packet throughput, as the experiment is arrival-limited.
The results are supported by the specific policy representation (per-channel tokens + QoS token) rather than attention alone, as the comparison baselines differ in multiple architectural components.
The study acknowledges limitations, including the use of synthetic AMC templates, ideal sensing, and simulation-only evaluation, precluding claims of causal attention mechanisms or deployed SDN performance.

The work positions itself as a step toward integrating physical-layer uncertainty and service-class differentiation into DRL-based spectrum access, offering a more robust solution for bursty IoT traffic scenarios.

Technical Summary: TACAN for Dynamic Spectrum Access

More like this