Hyperparameter Learning Optimization for Neural Receiver Models with Bayesian Networks

A comprehensive framework for hyperparameter optimization is developed in neural receiver models for Orthogonal Frequency Division Multiplexing (OFDM) wireless communication systems using the Estimation of Bayesian Network Algorithm (EBNA). The proposed methodology addresses the critical challenge of selecting optimal hyperparameters for convolutional neural network-based receivers in 5G OFDM systems. Through the application of EBNA, an algorithm that constructs probabilistic models to capture hyperparameter interdependencies, we perform a systematic optimization of both the learning rate and the number of residual blocks, with the objective of maximizing the Bit-Metric Decoding (BMD) rate. Our experimental results demonstrate that the optimized neural receiver achieves performance comparable to baseline models while requiring only 5% of the original training iterations, suggesting the existence of local optima that can be efficiently identified through Bayesian optimization. The hierarchical framework developed provides a scalable approach for future hyperparameter exploration in neural receiver communication systems.