Robust LSTM-Based Channel State Estimation for Next-Generation Wireless Communication Systems

Abstrak

Long Short-Term Memory (LSTM) networks conventionally employ the hyperbolic tangent (tanh) and sigmoid functions for state and gate activations, respectively. Although numerous alternative activation functions have been introduced for deep neural networks, their potential to improve LSTM performance in wireless communication tasks remains insufficiently explored. This paper investigates the replacement of the tanh state activation function in LSTM architectures with three alternatives: the Bi-tanh1, the Elliott, and the Gaussian Error Linear Unit (GELU). The resulting modified LSTM networks are evaluated as channel state estimators (CSEs) for 5G Orthogonal Frequency-Division Multiplexing (OFDM) systems operating over Rayleigh fading channels. Simulation results demonstrate that all proposed activation-function-based LSTM CSEs substantially outperform conventional Least Squares (LS) and Minimum Mean Square Error (MMSE) estimators, particularly under low pilot-density conditions. Among the evaluated configurations, the GELU-based LSTM CSE achieves the lowest symbol error rate (SER) across all tested signal-to-noise ratios (SNRs), establishing GELU as a strong candidate for robust channel estimation in next-generation wireless receivers

Identifier

• OpenAlex: https://openalex.org/W7166528349
• DOI: https://doi.org/10.66279/jdgvdt83