Enhancing LMMSE Performance with Modest Complexity Increase via Neural Network Equalizers
Vadim Rozenfeld, Dan Raphaeli, Oded Bialer · 2024
Read Enhancing LMMSE Performance with Modest Complexity Increase via Neural Network Equalizers by Vadim Rozenfeld, Dan Raphaeli, Oded Bialer (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2411-01517.txt.
Description
Abstract The BCJR algorithm is renowned for its optimal equalization, minimizing bit error rate (BER) over intersymbol interference (ISI) channels. However, its complexity grows exponentially with the channel memory, posing a significant computational burden. In contrast, the linear minimum mean square error (LMMSE) equalizer offers a notably simpler solution, albeit with reduced performance compared to the BCJR.
Catalog record
| CREATOR | Vadim Rozenfeld, Dan Raphaeli, Oded Bialer |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2411-01517
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Enhancing LMMSE Performance with Modest Complexity Increase via Neural Network Equalizers Vadim Rozenfeld, Dan Raphaeli, Oded Bialer Abstract The BCJR algorithm is renowned for its optimal equalization, minimizing bit error rate (BER) over intersymbol interference (ISI) channels. However, its complexity grows exponentially with the channel memory, posing a significant computational burden. In contrast, the linear minimum mean square error (LMMSE) equalizer offers a notably simpler solution, albeit with reduced performance compared to the BCJR. Recently, Neural Network (NN) based equalizers have emerged as promising alternatives. Trained to map observations to the original transmitted symbols, these NNs demonstrate performance similar to the BCJR algorithm. However, they often entail a high number of learnable parameters, resulting in complexities comparable to or even larger than the BCJR. This paper explores the potential of NN-based equalization with a reduced number of learnable parameters and low complexity. We introduce a NN equalizer with complexity comparable to LMMSE, surpassing LMMSE performance and achieving a modest performance gap from the BCJR equalizer. A significant challenge with NNs featuring a limited parameter count is their susceptibility to converging to local minima, leading to suboptimal performance. To address this challenge, we propose a novel NN equalizer architecture with a unique initialization approach based on LMMSE. This innovative method effectively overcomes optimization challenges and enhances LMMSE performance, applicable both with and without turbo decoding. I Introduction Reliable data transmission over an ISI channel is vital for modern communication systems, where multiple symbols interfere, distorting the received signal and complicating decoding. At the receiver, an equalizer is deployed to mitigate ISI effects and reconstruct the original signal. With known channel state information (CSI) to the receiver, equalizer and decoder can iteratively work together in a turbo equalization [ 1 ] manner, where feedback information from the decoder is looped back to the equalizer. The optimal equalization technique for BER minimization is nonlinear, e.g., BER-optimal BCJR algorithm [ 2 ] , maximizing the a posteriori probability (MAP) estimation of the data. However, full-state BCJR entails exponential computational complexity growth with channel memory. Reduced-state BCJR methods [ 3 , 4 , 5 ] address complexity concerns, but may exhibit inferior performance compared to full-state BCJR. Additionally, the complexity of reduced-state alternatives still grows exponentially with channel memory. An LMMSE equalizer [ 6 ] , provides a computationally efficient alternative, albeit with a trade-off in terms of BER performance. In turbo equalization, BCJR typically outperforms LMMSE due to its ability to generate more reliable log-likelihood ratio (LLR) outputs, regardless of the signal-to-noise ratio (SNR) and the number of iterations employed. Contrary to the model-based algorithms reviewed above, data-driven models, particularly NNs, are alternative approach for performing equalization tasks. Unlike traditional model-based algorithms, data-driven models are trained to directly map observations (data) to the originally transmitted symbols. In fact, NNs are capable to extract relevant features from observations themselves, and therefore a data-driven equalizer does not require to be explicitly fed with CSI, assumed trained on sufficiently large dataset. As an example, N-BCJR [ 7 ] is recurrent neural network (RNN) based decoder for sequential codes with close to optimal performance on the AWGN channel. The work [ 8 ] proposes a joint equalizer and decoder, each of which is implemented via NN, to achieve equalization and decoding processes without requiring knowledge of the CSI. The work [ 9 ] demonstrates blind equalization via variational auto-encoder (VAE). Examples of hybrid data-driven/model-driven receivers proposed for symbol detection tasks, both with and without CSI available to the receiver, include BCJRNet [ 10 ] and ViterbiNet [ 11 ] . While many of the aforementioned models demonstrate performance within a small gap of conventional algorithms, this often comes at the prohibitive cost of a high number of learnable parameters and operations carried out per received symbol per iteration. In fact, BCJRNet and ViterbiNet already have thousands of learnable parameters, far exceeding computational load of full-state BCJR. In MetNet [ 12 ] , the complexity of ViterbiNet is further reduced, although it remains three times higher than that of the full-state BCJR. Additionally, the models reviewed above are not trained to integrate seamlessly with standard decoders or to perform optimally in turbo equalization. In this paper, we investigate the performance potential of NN-based equalization using a limited number of learnable parameters. The benefit of such NN lies in its reduced computational complexity and fast running time. We design an NN equalizer with complexity comparable to LMMSE, which surpasses LMMSE and achieves a modest performance gap from the MAP equalizer. A significant challenge with NNs having a small parameter count is their susceptibility to converging to local minima, resulting in sub-optimal performance. To address this, we propose a novel NN equalizer architecture built upon LMMSE, which enhances its capabilities, applicable both with and without turbo decoding. This approach allows us to construct a compact NN that can be initialized with LMMSE to mitigate optimization issues. The main contributions of the paper are as follows: 1. Highlighting the difficulty in optimizing neural network equalizers with limited parameters, which tend to settle into sub-optimal local minima, failing to surpass LMMSE performance. 2. Introducing a NN-based equalizer design that addresses the optimization challenge and exceeds LMMSE performance while maintaining similar complexity levels. 3. Showcasing that NN-based equalizers offer a low-complexity solution for narrowing the BER gap between LMMSE and MAP, both with and without decoding. II System Model and Equalization Algorithms We consider a system with an information source producing independent and uniformly distributed (IUD) data bits \mathbf{b}\triangleq[b_{0},b_{1},\cdots,b_{K-1}] . For integrity reasons, a block is protected by LDPC code, yielding a code-word \mathbf{c}\triangleq[c_{0},c_{1},\ldots,c_{U-1}] . It is then partitioned into Q blocks, each with q=\log_{2}M bits, \mathbf{c}=[\mathbf{c}_{0},\ldots,\mathbf{c}_{Q-1}] , and mapped to M-ary PAM symbols \mathbf{x}\triangleq[x_{0},x_{1},\ldots,x_{Q-1}]^{T} using a binary Gray-code function. The AWGN ISI channel model is given by: z_{n}=h_{0}x_{n}+\sum\limits_{i=1}^{M_{h}}h_{i}x_{n-i}+w_{n},\quad n=0,1,\ldots,Q-1, (1) where the second term is ISI. We assume x_{n},w_{n}\text{ and }z_{n} are zero outside \{0,1,\ldots,Q-1\} . Here, \mathbf{h}=[h_{0},h_{1},\cdots,h_{M_{h}}] is the invariant and known to the receiver channel impulse response (CIR), \mathbf{z}=[z_{0},z_{1},\cdots,z_{Q-1}]^{T} be the receiver input and w_{n}\sim\mathcal{N}(0,\sigma^{2}_{w}) are noise samples, all are real-valued. On the receiver side a channel equalizer is operating to counteract the distortion caused by the ISI channel \mathbf{h} and a decoder is then recovering the original data. As both equalizer and decoder operate on same set of received data, it is quite beneficial to recover the data bits via turbo equalization, which involves iteratively exchanging information between the soft-in soft-out (SISO) equalizer and a SISO decoder. Namely, the equalizer associates an LLR estimation to code bit c_{n} using: L(c_{n})\triangleq\ln\left(\frac{P(c_{n}=0|\mathbf{z})}{P(c_{n}=1|\mathbf{z})}\right),\quad n=0,1,\ldots U-1, (2) or the corresponding soft-bit [ 15 ] : \Lambda(c_{n})=\tanh{(L(c_{n})/2)}. (3) While common choices for equalizers include MAP and LMMSE, our interest lies in an NN-based alternative. II-A LMMSE and MAP Equalizers LMMSE estimates \hat{x}_{n} by minimizing MSE \mathbb{E}(|x_{n}-\hat{x}_{n}|^{2}) . To this end, it processes a frame of observations with a length of N=N_{1}+N_{2}+1 at time n which described by: \mathbf{z}_{n}\triangleq[z_{n-N_{2}},\dots,z_{n},\dots,z_{n+N_{1}}]^{T} (4) to compute the estimation \hat{x}_{n} of x_{n} by: \hat{x}_{n}=\mathbb{E}(x_{n})+\mathrm{Cov}(x_{n},\mathbf{z}_{n})\mathrm{Cov}(\mathbf{z}_{n},\mathbf{z}_{n})^{-1}(\mathbf{z}_{n}-\mathbb{E}(\mathbf{z}_{n})). (5) For initial equalization step with no a priori information available, [ 6 ] suggests \mathbf{f}=[f_{N_{2}},\ldots,f_{0},\dots,f_{-N_{1}}]^{T} for each time step is set to: \mathbf{f}\triangleq(\sigma^{2}_{w}\mathbf{I}_{N}+\mathbf{HH^{T}})^{-1}\mathbf{h}_{n}, (6) with f_{0} as center tap of \mathbf{f} . Here, \mathbf{H} is N\times(N+M_{h})