Abstract
In wireless communications, the Multi-Access Channel (MAC) is a fundamental model where multiple transmitters communicate with a single receiver. Traditionally, interference between these users is managed through orthogonalization, such as Time-Division Multiple Access (TDMA). However, information theory suggests that the presence of feedback can enable transmitters to cooperate rather than merely avoid each other. This paper introduces MBAF (Multi-access Block Attention Feedback), a fully learnable channel code that utilizes Transformer architectures and an iterative decoding strategy. We demonstrate that MBAF allows independent transmitters to autonomously learn cooperative strategies, significantly outperforming state-of-the-art benchmarks and exceeding the theoretical capacity limits of channels without feedback in the finite block length regime.
Introduction
The challenge of the Multi-Access Channel (MAC) is often compared to a crowded room where multiple speakers attempt to address a single listener simultaneously. Standard protocols typically mitigate the resulting interference by assigning specific time slots or frequency bands to each user. While effective, this approach does not fully exploit the potential for coordination.
Information theory indicates that if a receiver can provide feedback to the transmitters, those transmitters can synchronize their efforts to achieve higher data rates. Despite this theoretical promise, designing practical codes that exploit feedback—especially for short message lengths—remains a significant hurdle. While deep learning has revolutionized point-to-point coding, its application to multi-user feedback scenarios is still nascent. This research proposes MBAF to bridge this gap, asking whether neural networks can teach autonomous agents to cooperate in a shared medium.
Methodology
The MBAF framework is built upon a fully learnable auto-encoder architecture, specifically optimized for the complexities of multi-user interference and feedback integration.
Transformer-Based Encoder
Unlike traditional recurrent neural networks (RNNs) that process sequences linearly, MBAF utilizes a Transformer backbone. The self-attention mechanism allows the encoder to evaluate the entire history of transmitted symbols and received feedback simultaneously. This global context is crucial for determining the optimal next symbol to minimize collective error.
Knowledge Vector and MIPSE
Each transmitter maintains a "knowledge vector" that encapsulates its original message, its transmission history, and the feedback received from the channel. The system employs MAC Iterative Parity Symbol Encoding (MIPSE), an algorithmic flow where transmitters generate symbols in discrete blocks, updating their internal knowledge state after every feedback cycle from the receiver.
Iterative Decoding (MJSPSD)
At the receiver side, the authors introduce MAC Joint Successive Parity Symbol Decoding (MJSPSD). Rather than a single-pass decoding attempt, the receiver generates an initial hypothesis, which is then fed back into the decoder as a "belief" signal. Over multiple iterations ($N_{iter}$), the decoder refines its prediction, significantly reducing the probability of error compared to standard one-shot decoders.
Training Strategy
To ensure convergence in complex multi-user environments, the model is trained using a curriculum learning approach. The system is initially exposed to "easy" scenarios with a high signal-to-noise ratio (SNR) and is gradually challenged with noisier environments as its performance stabilizes.
Experimental Results
The performance of MBAF was evaluated using a synthetic Additive White Gaussian Noise (AWGN) MAC with two transmitters, each sending a message of $K=51$ bits.
Performance Benchmarks
MBAF was compared against GBAF-TD, a state-of-the-art deep learning-based code adapted for TDMA. The results showed a substantial performance gap. At an SNR of 1 dB, MBAF achieved a Block Error Rate (BLER) of $1.7 \times 10^{-7}$, whereas the TDMA-based benchmark only reached a BLER of $1.8 \times 10^{-3}$. This represents a four-order-of-magnitude improvement in reliability.
Emergent Cooperation
One of the most significant findings is that the neural networks learned to cooperate without explicit instructions. By analyzing the feedback, transmitters were able to decode portions of each other's messages and coordinate their subsequent transmissions. This cooperative behavior allowed the system to achieve data rates that exceed the theoretical upper bound for a MAC channel without feedback, providing empirical evidence that the model successfully exploits the feedback link for coordination.
Discussion and Limitations
While MBAF represents a significant step forward in learnable multi-user coding, several challenges remain for real-world implementation:
- Feedback Reliability: The current model assumes "perfect" feedback, where transmitters receive the exact signal seen by the receiver. In practical 5G or 6G applications, the feedback link would be subject to its own noise and latency.
- Scalability: The study focused on a two-user scenario. Scaling the Transformer architecture to a "massive MAC" environment (e.g., hundreds of IoT devices) would involve significant computational complexity.
- Message Length: The optimization was performed for short blocks ($K=51$). Further research is required to determine if these cooperative gains persist for longer data streams.
Conclusion
MBAF demonstrates that deep learning can move beyond simple interference avoidance toward active multi-user cooperation. By leveraging the attention mechanism and iterative decoding, the system provides a robust framework for high-performance communication in feedback-enabled multi-access channels. Future work will focus on adapting these models for imperfect feedback and expanding the architecture to more complex topologies, such as broadcast channels.
Summary updated