Abstract
In this paper, we explore neural network-based strategies for performing symbol detection in a MIMO-OFDM system. Building on a reservoir computing (RC)-based approach towards symbol detection, we introduce a symmetric and decomposed binary decision neural network to take advantage of the structure knowledge inherent in the MIMO-OFDM system. To be specific, the binary decision neural network is added in the frequency domain utilizing the knowledge of the constellation. We show that the introduced symmetric neural network can decompose the original $M$-ary detection problem into a series of binary classification tasks, thus significantly reducing the neural network detector complexity while offering good generalization performance with limited training overhead. Numerical evaluations demonstrate that the introduced hybrid RC-binary decision detection framework performs close to maximum likelihood model-based symbol detection methods in terms of symbol error rate in the low SNR regime with imperfect channel state information (CSI).
Introduction
MIMO-OFDM is a cornerstone of modern 4G and 5G wireless networks. However, symbol detection at the receiver faces significant challenges. Traditional model-based methods, such as Maximum Likelihood (ML) detection, are computationally prohibitive (NP-hard) as antenna counts and modulation orders (e.g., 256-QAM) increase. Furthermore, real-world hardware non-linearities from power amplifiers and converters are difficult to capture with standard linear models.
While deep learning offers a potential solution, standard neural networks (NNs) typically require massive datasets to generalize. In practical 5G links, training data is scarce, with only about 10% of the signal dedicated to known "pilot" symbols. This "Data Deficit" makes generic NNs impractical for rapidly changing wireless channels.
Proposed Methodology: Structure-Aware Detection
The proposed approach integrates physical structural knowledge into a hybrid neural network architecture, avoiding the "black box" pitfalls of generic deep learning.
Reservoir Computing (RC)
The system utilizes Reservoir Computing, a type of Recurrent Neural Network (RNN) where internal connections are fixed and random, and only the output layer is trained. This architecture efficiently captures the time-domain dynamics of multipath fading with minimal training overhead.
Binary Decision Decomposition
To handle high-order modulation like 16-QAM, the authors decompose the $M$-ary detection problem into simpler binary classification tasks. By applying a shifting process, the network performs a sequence of binary decisions rather than a single complex classification.
Inductive Bias and Structural Knowledge
The network is initialized with the structure of the Inverse Discrete Fourier Transform (IDFT), aligning the model with the fundamental physics of OFDM signal generation. Additionally, weight sharing across sub-carriers exploits frequency-domain channel correlations, further reducing the learning burden.
Experimental Evaluation
The framework was evaluated using the WINNER II channel model (urban outdoor-to-indoor NLOS) in a $2 \times 2$ MIMO configuration with 256 sub-carriers and 16-QAM modulation. To reflect real-world constraints, training was limited to just two MIMO-OFDM symbols, matching the pilot structure of 3GPP standards.
Key Results
Numerical results indicate that the hybrid RC-binary framework achieves near-optimal performance. Under imperfect Channel State Information (CSI) using LMMSE channel estimation, the proposed method significantly outperforms standard linear detectors (LMMSE-SD) and approaches the performance of the ML detector in low SNR regimes.
The "knowledge of structure" approach allows the model to learn only the deviations from the ideal model (noise and non-linearities) rather than the entire physics of radio propagation, leading to superior generalization from limited data compared to pure Multi-Layer Perceptron (MLP) approaches.
Conclusion and Future Work
This research demonstrates that embedding domain-specific knowledge into neural network architectures can overcome the data deficit in wireless communications. Future work will explore scaling this approach to Massive MIMO systems and even higher modulation orders like 256-QAM. Additionally, hardware implementation on FPGAs is necessary to validate real-time processing latency and power efficiency in live environments.
Summary updated