Iterative Learning for Reliable Underwater Link Adaptation (Student Abstract)

Junghun Byun, Yong-Ho Cho, Tae-Ho Im, Hak-Lim Ko, Kyung-Seop Shin, Ohyun JoAAAI 2020 · DOI

Introduction

Underwater Wireless Sensor Networks (UWSNs) face unique challenges compared to terrestrial wireless systems. In typical terrestrial environments, the Signal-to-Noise Ratio (SNR) serves as a reliable predictor for the Bit Error Rate (BER). However, underwater channels are significantly more complex and chaotic, influenced by variables such as salinity, temperature, multipath delay, and Doppler shifts. Consequently, a high SNR does not inherently guarantee a reliable signal.

Traditional Adaptive Modulation and Coding (AMC) techniques, which select the transmission speed (MCS level) based on channel conditions, often fail in underwater settings because they rely too heavily on SNR. This dependency leads to either frequent connection failures or overly conservative, slow data rates. This research proposes a machine learning-based iterative framework designed to move beyond simple SNR metrics, utilizing multi-parameter models to achieve more robust and efficient link adaptation.

The Iterative Learning Framework

The core of the proposed solution is an iterative learning framework that employs two distinct machine learning models in a closed loop to optimize transmission parameters.

Experimental setup in the Yellow Sea (a) and the proposed iterative learning framework (b).
  1. MCS Prediction Model: This model predicts an initial Modulation and Coding Scheme (MCS) level based on the current underwater channel vector.
  2. BER Prediction Model: This model estimates the resulting Bit Error Rate (BER) for the specific MCS level selected by the first model.

The iterative nature of the framework is critical: the system evaluates whether the predicted BER meets a predefined reliability threshold. If the predicted error is too high, the system updates the MCS level and re-runs the prediction. This process continues until the system identifies the highest possible data rate that maintains a reliable connection. The researchers evaluated two architectures for these models: Multiple Linear Regression (MLR) and Multi-Layer Perceptron (MLP).

Experimental Dataset

Unlike many studies that rely on simulations, this research utilized a custom-built, real-world dataset collected in the Yellow Sea (Gulf of Incheon, South Korea). Data was gathered using OFDM-based acoustic communication testbeds. The resulting dataset contains 10,151 samples, each featuring a comprehensive "Channel Vector." This vector includes SNR, Carrier to Interference and Noise Ratio (CINR), delay spread, maximum delay component, Doppler frequency shift, and both coded and uncoded BER.

Performance Evaluation and Results

The field tests confirmed that SNR alone is an insufficient predictor of performance in underwater environments, justifying the shift toward multi-parameter machine learning models. The results demonstrate that the iterative approach allows the system to safely "push the envelope" of data rates compared to conventional, conservative models.

Comparison of predicted throughput across SNR-based, MLR, and MLP models at different measured BER levels.

Key findings from the evaluation include:

  • Throughput Improvement: The iterative learning model achieved up to a 25% increase in total throughput compared to conventional SNR-based benchmarks.
  • Model Accuracy: The MLP model (a type of neural network) outperformed the simpler MLR model, indicating that the relationship between underwater channel parameters and link performance is non-linear and highly complex.
  • Reliability: The MLP model successfully predicted higher data rates while consistently satisfying target BER requirements, whereas the SNR-based model frequently underestimated the channel's actual capacity.

Conclusion and Future Work

While the MLP model provided the most accurate results, it introduces higher computational complexity—a significant factor for battery-powered underwater sensors. Future research will focus on balancing model complexity with energy efficiency. Additionally, the authors plan to incorporate direct environmental variables such as water temperature, currents, and salinity into the models to further refine predictions. Finally, as the current data is specific to the Yellow Sea, further testing is required to determine how well these models generalize to different oceanic environments, such as deep-sea or varying coastal regions.

Summary updated

BibTeX
@inproceedings{byun2020iterative,
  author = {Junghun Byun and Yong-Ho Cho and Tae-Ho Im and Hak-Lim Ko and Kyung-Seop Shin and Ohyun Jo},
  title = {{Iterative Learning for Reliable Underwater Link Adaptation (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7152},
}

Loading related content...

Figure