Sliding Bidirectional Recurrent Neural Networks for Sequence Detection in Communication Systems

Nariman Farsad, Andrea GoldsmithICASSP 2018 · DOI

TL;DR

This paper introduces a novel deep learning approach, the Sliding Bidirectional Recurrent Neural Network (SBRNN), for detecting sequences of transmitted symbols in communication systems where the underlying channel characteristics are unknown. By training on experimental data from a molecular communication platform, the SBRNN demonstrates superior performance compared to existing methods, paving the way for more robust communication systems in challenging environments.

What Problem Does This Paper Address?

Designing and analyzing communication systems traditionally relies on accurate mathematical models of the communication channel. However, in emerging areas like molecular communication (using chemical signals for information transfer) or in complex environments like underwater acoustics, these channel models are often unknown, difficult to derive, or analytically intractable. This lack of a model poses a significant challenge for developing effective detection algorithms, which are crucial for estimating transmitted symbols from noisy received signals. Existing machine learning approaches often still assume some knowledge of the channel model. This paper tackles the problem of designing robust detection algorithms for communication systems when the channel model is completely unknown, focusing on the critical task of sequence detection.

Methods

The core contribution is the Sliding Bidirectional Recurrent Neural Network (SBRNN) detector. Traditional Recurrent Neural Networks (RNNs), like LSTMs, are good for sequence estimation but are typically feed-forward, meaning they don't consider future observations when estimating a past symbol. Bidirectional RNNs (BRNNs) address this by processing the sequence in both forward and backward directions, but they require processing entire blocks of data, which becomes computationally infeasible for long sequences.

The SBRNN overcomes this by:

  1. Fixing a BRNN length (L): This length can be treated as a hyperparameter.
  2. Sliding the BRNN window: Instead of processing data in fixed, non-overlapping blocks, the BRNN window slides one symbol at a time across the incoming data stream.
  3. Averaging estimates: For each symbol, the SBRNN averages the detection outputs from all BRNN windows that overlap with that symbol. This dynamic programming-like approach allows for real-time detection as new symbols arrive, while still leveraging the benefits of bidirectional processing.

The paper also explores other deep learning architectures for comparison, including fully connected networks (Base-Net), 1D Convolutional Neural Networks (CNN-Net), and standard LSTM and BLSTM networks for sequence detection.

Datasets

The authors utilized an experimental platform for in-vessel molecular communication to collect data. This platform uses chemical signals (acids and bases) to encode information, and the received signal is measured as changes in pH level. Data was collected for four different symbol durations (250 ms, 334 ms, 380 ms, and 500 ms), corresponding to data rates of 4, 3, 2.6, and 2 bits per second (bps).

For each symbol duration, random bit sequences of length 120 were transmitted 100 times. A total of 12k bits per symbol duration were generated. The dataset was split into training (84 transmissions per symbol duration, totaling 40,320 bits) and testing (16 transmissions per symbol duration, totaling 7,680 bits). This experimental dataset is novel and specifically designed for evaluating communication systems where channel models are unknown.

Key Insights and Ideas

  • Deep Learning as a Model-Free Solution: The paper powerfully demonstrates that deep learning can effectively train communication system components, like detectors, without relying on explicit mathematical models of the channel. This is a paradigm shift for systems where modeling is a bottleneck.
  • Sequence Detection is Crucial: The results highlight that sequence detection methods (RNNs, BRNNs) significantly outperform symbol-by-symbol detectors, especially in the presence of Inter-Symbol Interference (ISI), which is prevalent in chemical communication.
  • The SBRNN's Real-Time Advantage: The SBRNN's sliding window approach is key to enabling real-time, dynamic detection. It avoids the computational burden of re-processing entire blocks and allows for immediate symbol estimation as new data arrives.
  • Molecular Communication as a Testbed: The use of a real-world molecular communication platform provides a challenging and realistic environment to validate the proposed techniques, showcasing their applicability beyond simulated scenarios.

Key Results

The SBRNN detector achieved the best Bit Error Rate (BER) performance across various transmission rates. For instance, at a 380 ms symbol duration (2.6 bps), the SBRNN achieved a BER of 0.0005, significantly outperforming the baseline detector (0.0797) and other deep learning methods. The paper also notes that the SBRNN is resilient to changing channel conditions and can perform detection without explicit channel state information. The authors also mention achieving a data rate of 2 bps for a text messaging service using their trained detectors, which is an order of magnitude higher than previous experimental systems.

Limitations and Future Work

  • Dataset Size: The authors acknowledge that the experimental dataset is not very large due to the laborious nature of collecting measurements. They suggest that training with larger datasets would be beneficial and is explored in their future work.
  • Hyperparameter Tuning: While grid search was used for hyperparameter tuning, optimizing this process further could yield even better results.
  • Generalizability: While the SBRNN is shown to be robust, further investigation into its performance across a wider range of communication environments and channel complexities would be valuable.
  • Future Work: The paper mentions a journal extension where they explore a Poisson channel model for non-reactive molecular communication, showing the SBRNN's performance can approach that of an optimal Viterbi detector in low noise. They also emphasize the potential of SBRNN for future communication systems where traditional modeling is difficult.

Summary updated

BibTeX
@inproceedings{farsad2018sliding,
  author = {Nariman Farsad and Andrea Goldsmith},
  title = {{Sliding Bidirectional Recurrent Neural Networks for Sequence Detection in Communication Systems}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2018},
  doi = {10.1109/icassp.2018.8462140},
}
Figure