Efficient implementation and optimization of adVanced messAge paSsing-based receIvers fOr future wireless Networks – EVASION

The project started from the available state of the art knowledge about Expectation Propagation (EP), Belied Propagation (BP), Approximate Message Passing (AMP) and other message passing techniques. The first task was to identify the most mature and promising use cases for the use of message passing / EP processing, namely equalization and MIMO/NOMA detection. One of the main objectives of WP1 and WP2 was then to design low-cost and robust variants of the EP receivers.
Concerning equalization, the first step was to analyze the complexity of the Frequency Domain Linear Equalizer with Expectation Propagation. This evaluation revealed that soft mapping and soft demapping operations were the main contributors to the overall receiver complexity, whereas the complexity of the equalizer part was not critical. Starting from this observation, multiple simplifications have been studied and compared, and some of then can be linked to sound theoretical analysis and are not just linked to heuristics. The first simplified soft demappers were conceived without a priori from the decoder. Then a general version taking into account the a priori information from the decoder was also proposed.
Concerning MIMO/NOMA detection, the project investigations determined that, in particular for massive MIMO schemes, the detector complexity is the main contributor to the receiver complexity. Hence the effort has been concentrated on studying simplified EP detection algorithm. Nevertheless, the work done in simplifying the calculation of the EP feedback could be largely reused for the MIMO case too. For lowering the receiver complexity, many works in the open literature proposes to approxime the most costly operation, i.e. matrix inversion for vector EP. Our approach here is different, the effort was concentrated on directly working on the factor graph, which allowed to obtain simpler and efficient scalar EP algorithms.
Given the simplifications and optimizations on the studied detectors and demodulators provided by WP1/WP2, a candidate algorithm has been agreed upon by the consortium. The choice was the frequency domain equalizer based on EP, since at the selection moment, it had a higher technical maturity. This receiver has been implemented in py-AFF3CT, which is a Python wrapper for the open-source platform AFF3CT. Then the algorithm has been verified with the help of the MATLAB simulator. From the validated py-AFF3CT code, a fixed-point implementation has been generated (see D3.3) and the impact on performance has been evaluated (see D3.1). In parallel the study of the hardware (FPGA) architecture has been carried out. The VHDL descriptions of the proposed fixed-point architecture for the FPGA has been produced and integrated in a hardware-on-the-loop demonstrator. This methodology was effective in solving the complexity issue of the EP receiver, since a very interesting implementation was obtained thanks to the joint work of all the partners.

For equalization, the focus was on simplifying the EP soft demapper/mapper. This goal was achieved: two simplified structures were proposed for the case without soft information from the decoder, and one general structure for the case with a priori information from the decoder, which was then selected for implementation in FPGA. The algorithmic simplifications brought up to a factor of 10 of complexity reduction for high order constellations, and performance degradation is in general very limited.
1 international journal article and 2 international conference papers have been published on this subject. 4 patents have been filed in France and are being extended to Europe.

For MIMO detection, different proposals using BP and EP where evaluated. After comparing them, the most promising simplified EP-based MIMO detection, in terms of performance/complexity trade-off belons to the class of scalar EP algorithms. It combines a precoding processing of the channel matrix and received signal with an optimized scheduling of the messages. This receiver allows to reduce the computational complexity by a factor of 2 for massive deployments with respect to vector EP. Moreover, in terms of performance for the first turbo iterations (which is the interesting case in practice), the proposed receiver sticks to the vector EP performance. A fixed-point version was developed which showed even greater robustness in very selective test channels. The fixed point code was then used for implementing the soft EP mapper/demapper in the FPGA of a Xilinx ZYNQ Z2 evaluation board, where only a few percent of the total resources of the FPGA were used.
This work has led to the publication of 1 international journal article, 3 international conference papers, and a 2 national conference papers.

EVASION has worked a lot on the application of deep unfolding to iterative receiver based on EP or AMP. The optimisation of damping parameters between the soft demapper and the decoder showed how extrinsic information from the decoder is not always optimal and an opportune mix of a posteriori information should be used in certains cases. Moreover, Gated Recurrent Unit (GRU), whose structure was optimized by injecting expert knowledge, was also proposed to generate damping parameters at the layers of the unfolded structure. This receiver improved the performance over non-IA-aided receivers, and also the interpretability of the GRU.
The use of Graph Neural Networks (GNNs) to estimate a posteriori symbol distribution for iterative receivers was also investigated. We showed that in some cases, the GNN structure can be significantly optimized thus reducing the complexity of the neural network with limited performance degradation.
This work has led to the publication of 3 international conference papers, and a journal article is in preparation.

EVASION achieved a substantial improvement of the technological maturity of Expectation Propagation (EP) processing.
For the frequency domain equalization functionality (which is widely used in receivers for Single Carrier or SC-FDMA communication systems), the EVASION project succeeded in proposing algorithmic simplifications which drastically reduced the complexity of the EP soft demapper, a functional block whose implementation is not yet assessed. Moreover, a fixed point version of the EP soft mapper was developed on the PyAff3ct platform, which was then used as a basis for a FPGA implementation, integrated in a hardware-on-the-loop demonstrator. A flexible and pipelined architecture has been obtained too. The resource consumption of the soft demapper on the FPGA of a Xilinx PYNQ Z2 board is only few percentage points of the total available resources, thus showing the effectiveness of the implementation.
Advances on the MIMO/NOMA detection schemes based on EP have also been produced. Namely, flexible scalar EP MIMO detector with QR pre-processing and optimized scheduling of the messages has been proposed and its performance assessed. This detector can be applied on several system configurations ranging from classical MIMO systems with low number of antennas, up to Massive MIMO systems with tens of antennas, always with small or negligible performance loss with respect to the initial and more complex vector EP detector. While the vector EP detector is too complex for Massive MIMO applications, the proposed receiver shows up to 50% reduction in the computational complexity (with one self-iteration, more with more self-iterations). Moreover, the low-complexity EP soft demapper implemented in FPGA for the equalization use case can also be reused here to ease the implementation. An interesting perspective consists in evaluating its performance when applied to the proposed MIMO detection scheme, and other solutions of the literature. There will also be room for further optimization of the low-complexity soft mapper for these MIMO/NOMA detectors.
The EVASION project produced also a lot of other original results especially in the field of Artificial Intelligence-aided receivers for MIMO and NOMA detectors based on EP or Approximate Message Passing, showing encouraging performance gains and opening exciting research paths.

4 dépôt de brevets
2 international journal papers
8 international conference papers
2 national conference papers
2 workshops ont he open source plateform Aff3CT
1 GdR IASIS workshop organised by the project
1 tutoriel at the international conférence EUSIPCO 2024

Recently, the academic and even the industrial community has showed a renewed interest in approximate Bayesian inference and derivative message passing techniques, thanks to new interference cancellation receivers based on expectation propagation (EP). These emerging solutions have shown near optimal performance with attractive complexity-performance trade-offs with respect to conventional approaches based, for instance, on belief propagation (BP). These new algorithms are also able to produce useful signal estimates uncorrelated with the observations, a property which considerably reduces error propagation in the interference cancellation process inherent to these iterative receivers. This fact makes them very promising candidates in all situations in which interference cancellation is used and/or necessary.
EVASION aims to extend EP-based message passing algorithms to a broad range of use cases for 4G/5G wireless cellular networks, mobile ad hoc networks and also high frequency (HF) communication systems. These use cases include for instance equalization for orthogonal frequency division multiplexing (OFDM) or single carrier (SC) systems and multi-dimensional constellation detection used in multi-user or single-user multiple input multiple output (MIMO) communications, or in non-orthogonal multiple access (NOMA) both for OFDM and SC, the latter being less investigated in the open literature. Comparisons with other related approximate message passing algorithms will be performed. Moreover, in certain NOMA schemes, multi-dimensional constellations can be indeed decoded using the message passing approach in a non-binary framework. In this context, we will also investigate the design of efficient channel coding schemes, reaching best achievable information rates when considering joint detection and decoding, and compare it with EP-based receiver with serial interference cancellation which has already shown interesting potential.
The project will improve EP-based message passing algorithm technological maturity in the previously quoted applications by investigating: (a) the role of imperfect/quantized channel state information, (b) the impact of variables’ quantization and dynamics restrictions, and (c) the interactions with channel coding. The consortium will select and publish part of the investigated receiver algorithms in AFF3CT, an open-source software toolbox.
We believe that such message passing algorithms have a common structure which allows finding generic implementation architectures, which can be adapted with limited changes to the variety of previously quoted applications. EVASION will study two types of architecture: (i) the classical iterative detection architecture based on a loopy factor graph, and (ii) the deep unfolded/unrolled architecture which is by essence a pipelined architecture. We recall that deep unfolding is an emerging paradigm in deep learning, which will be used to optimize both algorithm performance and its implementation.
Finally, the consortium will provide an FPGA implementation of selected message passing algorithms, under implementation constraints provided by the industrial partner. The concept of “hardware in the loop” that associates software simulations and hardware prototyping will be applied during the implementation of algorithms selected in the early phase of the project. This methodology enables to accelerate the validation of the architecture and, at the same time, to evaluate the architecture design impact on the transmission system performance.
EVASION is expected to produce significant innovation in the area of advanced receivers with EP and in particular of simplifications and optimizations for their practical implementation. Besides traditional scientific dissemination activities, the consortium will establish at the beginning of the project an agreed and common strategy for open source software production and intellectual property protection.