TL;DR
This paper introduces ViterbiNet, a novel deep learning approach that integrates neural networks into the Viterbi algorithm for symbol detection in digital receivers. ViterbiNet achieves near-optimal performance without requiring channel state information (CSI), making it robust to CSI uncertainty and applicable to complex channels where traditional Viterbi is challenging to implement.
What Problem Does This Paper Address?
Symbol detection is a crucial step in digital receivers, aiming to accurately recover transmitted symbols from noisy channel outputs. Conventional methods, like the Viterbi algorithm, are highly effective but rely on complete knowledge of the channel's statistical properties (CSI). Obtaining accurate CSI can be challenging, especially in dynamic or complex communication environments, leading to performance degradation or increased overhead for estimation. This paper addresses the need for a symbol detection method that can achieve high performance without explicit CSI, while also being adaptable to changing channel conditions.
Methods
The core innovation is ViterbiNet, which modifies the Viterbi algorithm by replacing its channel-model-dependent computations (specifically, the log-likelihood calculation) with a deep neural network (DNN). This DNN is trained to learn the mapping from channel outputs to log-likelihoods, effectively learning the channel's behavior from data rather than relying on a predefined model.
Key methodological aspects include:
- Deep Unfolding Inspiration: The design is inspired by deep unfolding techniques, where model-based algorithms are translated into neural network architectures. However, ViterbiNet integrates ML into the Viterbi algorithm rather than replacing it entirely.
- ML-based Log-Likelihood Computation: A DNN is trained to estimate the conditional probability $P(Y[i]|S[i-l+1], \dots, S[i])$ (channel output given transmitted symbols), which is then used to compute the log-likelihoods required by the Viterbi algorithm. This is achieved by estimating $P(S[i-l+1] | Y[i])$ and $P(Y[i])$ and using Bayes' theorem.
- Meta-Learning for Online Adaptation: To handle time-varying channels, a meta-learning approach is proposed. This allows ViterbiNet to adapt to new channel conditions in real-time using data from decoded blocks, without requiring new, explicit training datasets for each change. This is facilitated by using forward error correction (FEC) codes to generate reliable training signals from decoded blocks.
Datasets
The paper relies on simulated data generated from defined channel models (ISI with AWGN, Poisson channel) for training and evaluation. No specific publicly available datasets are mentioned. The authors generate training samples by simulating channel transmissions and then use these to train ViterbiNet.
Key Insights and Ideas
- Decoupling Detection from CSI: The most significant insight is that high-performance symbol detection can be achieved by learning the channel's behavior from data, thereby bypassing the need for explicit CSI.
- ViterbiNet's Robustness: ViterbiNet demonstrates remarkable resilience to CSI uncertainty, outperforming traditional Viterbi algorithms when CSI is imperfect.
- Adaptability to Complex Channels: The data-driven nature of ViterbiNet makes it suitable for complex channel models (e.g., those with alpha-stable noise) where analytical modeling and CSI estimation are extremely difficult.
- Efficient Online Learning: The meta-learning approach enables ViterbiNet to track dynamic channel variations with minimal overhead, leveraging the structure of coded communication signals.
- Small Training Data Requirement: ViterbiNet can be trained effectively with relatively small datasets, making it practical for real-world applications.
Key Results
- Performance Approaching Optimal: ViterbiNet achieves symbol error rates (SER) very close to those of the CSI-based Viterbi algorithm, especially when trained with sufficient data.
- Outperforming ML Baselines: ViterbiNet outperforms previously proposed deep symbol detectors (like SBRNN) in scenarios with limited training data and CSI uncertainty.
- Resilience to CSI Uncertainty: In the presence of imperfect CSI, ViterbiNet shows significantly better performance than the Viterbi algorithm. For example, in AWGN channels, ViterbiNet with CSI uncertainty performs much better than Viterbi with CSI uncertainty.
- Effective Online Tracking: For block-fading channels, ViterbiNet with online training approaches the performance of CSI-based Viterbi, demonstrating its ability to adapt to changing channel conditions.
- Complex Channel Performance: ViterbiNet performs well in channels with alpha-stable noise, where traditional Viterbi is computationally prohibitive due to the difficulty in approximating PDFs.
Limitations and Future Work
- Computational Complexity Growth: Similar to the Viterbi algorithm, ViterbiNet's complexity is expected to grow exponentially with constellation size ($m$) and channel memory ($l$). The authors suggest that techniques like beam search could inspire future modifications to mitigate this.
- Sensitivity to Decision Errors in Online Training: The meta-learning approach for online training can be sensitive to errors in symbol detection, especially at low SNRs. The paper proposes using error detection codes and a threshold to mitigate this.
- Unknown Memory Length: The current work assumes the channel memory length ($l$) is known. Extending the decoder to handle unknown memory length is suggested as future work.
- Exploiting Other Signal Structures: The meta-learning approach currently leverages channel coding structure. Future work could explore using other protocol-induced structures like preamble sequences or pilots for meta-training.
- Generalization to Other Algorithms: The approach of integrating ML into model-based algorithms could be extended to other communication and signal processing algorithms (e.g., BCJR algorithm).
Summary updated