A Markov Chain Monte Carlo Detection Method Based on Deep Unfolding
Through the deep-expanded Markov chain Monte Carlo detection method, the gradient descent and perturbation noise parameters are optimized using deep learning technology, solving the problems of insufficient performance and high complexity in MIMO detection, and achieving an efficient detection strategy, which significantly reduces the Calculate overhead and detect delays.
Patent Information
- Application Number
- CN202211613281.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-15
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-12-15
AI Technical Summary
The existing MIMO detection technology is difficult to achieve efficient detection due to the performance-complexity compromise when expanding the system dimension. Bayesian inference technology has problems such as insufficient approximate accuracy and inefficient empirical parameter configuration.
Using the Markov chain Monte Carlo detection method based on deep expansion, the Markov chain Monte Carlo detection network is constructed, and the parameters of gradient descent and perturbation noise are optimized, combined with parallel sampling technology, the data model is built with dual-driven Detection strategy.
It significantly improves detection efficiency, reduces calculation complexity and detection delay, overcomes the complex line search process in the prior art, and achieves excellent convergence performance and efficient detection results.
Smart Images

Figure CN116015547B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a Markov chain Monte Carlo detection method based on deep expansion, and belongs to the technical field of wireless communications. Background Art
[0002] With the increasing data traffic and business demands, the design of future wireless communication systems increasingly needs efficient technologies and architectures. Multiple-Input Multiple-Output (MIMO) technology, as a key enabling technology inherent in the sixth generation of mobile communications, can greatly improve the bandwidth efficiency of the system to meet the requirements of diversified multimedia services for service quality, latency, performance and other indicators. However, the gains brought by MIMO technology are subject to the design of the detection strategy at the receiving end. In particular, considering the continuous expansion of the system dimension, the performance-complexity trade-off has become the core demand of the MIMO detection solution. Although existing technologies such as spherical decoding can achieve performance close to maximum likelihood detection, the hardware complexity makes it difficult to be practical in systems with medium-sized antenna configurations.
[0003] In response to the above challenges, detection technology based on Bayesian inference has attracted widespread attention due to its excellent performance and moderate complexity. Bayesian approximate inference can be divided into two paradigms: deterministic approximation and stochastic approximation. Deterministic approximation usually decomposes the target probability distribution in a specific form, or parameterizes it into a simple distribution. When it does not meet the model assumptions, it has the defect of insufficient approximation accuracy. Stochastic approximation completes statistical inference in a sampling manner. The most typical one is Markov chain Monte Carlo technology. Unlike deterministic approximation, when computing resources are sufficient, stochastic approximation represented by Markov chain Monte Carlo can approximate the exact solution to the inference problem. However, stochastic approximation also has problems such as low search efficiency and parameter configuration that is too empirical or model-dependent.
[0004] In recent years, with the widespread application of deep learning technology in the physical layer of wireless communications, combining deep learning's excellent feature extraction and inference capabilities to make up for the shortcomings of traditional modeling optimization methods has become an important means to break through the bottleneck of physical layer design. How to apply deep learning technology to the design of MIMO detection solutions, overcome the challenges of insufficient approximate accuracy and reliance on empirical parameter configuration in existing Bayesian inference technology, and build an efficient detection strategy driven by both data and models is urgently needed. Summary of the invention
[0005] Technical problem: In order to overcome the deficiencies in the prior art, the present invention provides a Markov chain Monte Carlo detection method based on deep expansion, aiming to utilize deep learning technology, especially a data model dual-drive technology based on deep expansion, to improve the problems of insufficient performance and inefficient empirical parameter setting in the prior art.
[0006] Technical solution: The technical solution adopted by the Markov chain Monte Carlo detection method based on deep expansion of the present invention specifically includes the following steps:
[0007] S1. Collect a training data set of the MIMO wireless communication system, where each piece of training data includes a channel matrix and noise variance of a channel realization, and a corresponding received signal vector;
[0008] S2, construct a deeply expanded Markov chain Monte Carlo detection network;
[0009] S3, using the collected training data set to train a deeply expanded Markov chain Monte Carlo detection network;
[0010] S4. In the deployment phase, the trained deep expanded Markov chain Monte Carlo detection network is used to obtain a sample list by parallel sampling. A decision is made based on the sample list to obtain the sample with the smallest residual norm in the sample list or the log-likelihood ratio as the detection result.
[0011] Furthermore, the deeply expanded Markov chain Monte Carlo detection network includes N layers of sub-sampling networks with the same structure, wherein the input of each layer of sub-sampling network includes: channel matrix, noise variance and received signal vector, as well as samples, sample residuals and sample residual norms output by the previous layer of sub-sampling network; the output of each layer of sub-sampling network includes: samples, sample residuals and sample residual norms of the current layer.
[0012] Furthermore, each layer of the sub-sampling network of the deeply expanded Markov chain Monte Carlo detection network specifically includes the following steps:
[0013] S201, executing a gradient descent process to calculate a mean of a proposal vector, wherein the step size of the gradient descent is trainable and is obtained by training a deeply expanded Markov chain Monte Carlo detection network;
[0014] S202, adding disturbance noise to determine the covariance of the proposal vector, wherein the amplitude of the disturbance noise is trainable and is obtained by training a deeply expanded Markov chain Monte Carlo detection network;
[0015] S203, projecting the elements of each dimension of the proposal vector to the transmission constellation point with the closest Euclidean distance to it, to obtain candidate samples;
[0016] S204, the ratio of the likelihood probability of the candidate sample to the likelihood probability of the output sample of the previous layer of sub-sampling network is used as the acceptance probability P acc ; Get a random number in, Indicates uniform distribution; judge Pacc Greater than or equal to P uni Is it true? If not, let the samples of the current layer sub-sampling network be the samples output by the previous layer sub-sampling network; if so, let the samples of the current layer sub-sampling network be candidate samples to obtain output samples; calculate the output sample residual and sample residual norm.
[0017] Furthermore, the gradient descent process described in step S201 uses a first-order acceleration method or a second-order Newton method; when using the first-order acceleration method, the gradient descent process includes multiple steps, and at the end of each step, the elements of each dimension of the current step estimation vector are projected to the transmission constellation point with the closest Euclidean distance to it to obtain additional samples.
[0018] Furthermore, the training method of the deeply expanded Markov chain Monte Carlo detection network specifically includes:
[0019] The channel matrix, noise variance and received signal are used as input features of the network, and the steps S201-S204 are completed through each layer of sub-sampling network, and samples of the elements of each dimension projected to the transmission constellation points are output to form a sample list;
[0020] Optimize the trainable parameters in the network in an unsupervised learning manner, including the step size of gradient descent and the amplitude of perturbation noise, and the loss function of training is the residual norm, expressed as:
[0021]
[0022] Where y represents the received signal, H represents the channel matrix, σ 2 represents the noise variance, represents the mathematical expectation of a small batch of received signals, channel matrices, and noise variances for the training data set, ||·|| represents the norm of the vector, is the sample in the sample list that minimizes the residual norm.
[0023] Furthermore, the method of projecting to the transmission constellation point uses a quantization function to act on the real part and the imaginary part of the complex value element respectively, specifically including:
[0024] The approximate quantization function Γ is used in training, and its expression is:
[0025]
[0026] Where x is the real continuous value parameter of the function input, M is the cardinality of the real value transmission constellation, m is the sequence number of the constellation symbol, tanh[·] is the hyperbolic function, η is the preset fitting coefficient, and E s is the average energy of the transmission constellation;
[0027] In the deployment phase, the precise quantization function Λ is used, which is expressed as:
[0028]
[0029] Here, ε(·) is a step function.
[0030] Furthermore, the parallel sampling method described in step S4 is specifically: using multiple Markov chain Monte Carlo detection networks with the same structure and parameters and with the same depth to generate samples for the same channel to form a sample list.
[0031] Beneficial effects: Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:
[0032] The present invention uses an efficient Markov chain Monte Carlo sampler that integrates gradient descent as the prototype of the detection method, expands it into a detection network driven by both data and models, optimizes the step size of the gradient descent part of the network and the amplitude of the disturbance noise through deep learning technology, improves the empirical parameter design in the prior art, and thus significantly improves the detection efficiency; in terms of computational complexity, the present invention avoids the complex line search process in the prior art, and at the same time, with the help of deep expansion tuning, achieves excellent convergence performance, thereby achieving performance comparable to that of the prior art with fewer sampling times, greatly reducing computational overhead and detection delay. In general, the present invention provides a detection solution that achieves a good compromise between performance and computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is a flow chart of an embodiment of the present invention;
[0034] Figure 2 is a structural block diagram of an embodiment of the present invention;
[0035] Figure 3 A comparison chart of the bit error rate (BER) performance of the embodiment of the present invention and the comparative method when the first-order acceleration method is used for the gradient descent part;
[0036] Figure 4 A comparison diagram of the bit error rate (BER) performance of the embodiment of the present invention and the comparative method when the second-order Newton method is used for the gradient descent part. DETAILED DESCRIPTION
[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0038] 1. System Model of This Embodiment
[0039] Consider a system with N tTransmitting antennas and N r MIMO system with 100 receiving antennas. In this embodiment, the number of antennas is N. t =N r =8. The information bit stream is subjected to constellation modulation and spatial data stream mapping to obtain a spatially multiplexed user data stream, which constitutes a transmission symbol vector x. In this embodiment, constellation modulation uses quadrature amplitude modulation (QAM). Specifically, the transmission constellation is a 16-QAM square constellation, and assume that the elements of x are The medium probability selection corresponds to the average transmission constellation energy E s =10, the real value transmission constellation is {-3, -1, +1, +3}, the cardinality is M = 4, and the actual transmission constellation is normalized according to the average energy. The received signal vector y can be expressed as:
[0040] y=Hx+v
[0041] Where H is N r ×N t dimensional complex-valued channel matrix. In this embodiment, the energy of each column of H is normalized, and v is a matrix with zero mean and variance σ 2 N r The detection problem addressed by this embodiment is that, given a triple {y, H, σ 2}, the transmitted symbol vector x is recovered. The optimal solution to this detection problem is given by the maximum likelihood solution, which can be expressed as:
[0042]
[0043] Among them, the objective function f(x) = ||y-Hx|| 2 , ||·|| represents the norm of the vector, argmin represents the minimum point of the objective function, and the constraints are Consider relaxing the constraints to the full N t dimensional complex space The least square (LS) solution to the detection problem can be obtained, which can be expressed as:
[0044]
[0045] Although the LS solution is suboptimal, relaxing the problem into a continuous space allows the application of gradient descent techniques for solving convex optimization problems, which can provide a rough direction for the optimal solution for the Markov Chain Monte Carlo-based sampling technique, thereby improving the search efficiency.
[0046] 2. Specific steps of this embodiment
[0047] like Figure 1 As shown, this embodiment provides a flowchart of a Markov chain Monte Carlo detection method based on deep expansion, including the following steps:
[0048] S1. Collect a training data set of a MIMO wireless communication system, where each piece of training data in the training data set includes a channel matrix and a noise variance of a channel realization, and a corresponding received signal vector;
[0049] S2, construct a deeply expanded Markov chain Monte Carlo detection network;
[0050] S3, using the training data set collected in step S1, training the deep expanded Markov chain Monte Carlo detection network constructed in step S2;
[0051] S4. In the deployment phase, the deeply expanded Markov chain Monte Carlo detection network trained in step S3 is used to obtain a sample list by a parallel sampling method; a judgment is made based on the sample list, and a sample with the smallest residual norm in the sample list or a log-likelihood ratio is obtained as a detection result.
[0052] Furthermore, step S2 expands the Markov chain Monte Carlo detection algorithm to obtain Figure 2 The structure diagram of the deeply expanded Markov chain Monte Carlo detection network shown in FIG. 1 includes N layers of sub-sampling networks with the same structure, where the input of the n-th layer of sub-sampling network includes the channel matrix H and the noise variance σ 2 and the received signal vector y, and the sample x output by the (n-1)th layer subsampling network n-1 , sample residual r n-1 =y-Hx n-1 and the sample residual norm l n-1 =||r n-1 ||, in particular, the initial sample x0 input to the first layer of sub-sampling network can be composed of randomly generated transmission constellation points, or the initial estimation result can be given by other detection techniques, and then projected to the transmission constellation point with the closest Euclidean distance; the output of the nth layer of sub-sampling network includes: the sample x0 of the current layer n , sample residual r n =y-Hx n and the sample residual norm l n =||r n ||, where n = 1, 2, ..., N is the sub-sampling network number. The following describes the working steps of each sub-sampling network layer by describing the n-th sub-sampling network layer:
[0053] S201, executing a gradient descent process to calculate a mean of a proposal vector. In an optional implementation of this embodiment, the gradient descent process includes a first-order acceleration method or a second-order Newton method;
[0054] Specifically, when using the first-order acceleration method, the gradient descent process includes N g Step, where N g is the total number of steps, each step includes:
[0055] Taking the estimated vector of the previous step as the starting point, the momentum update is performed, which can be expressed as:
[0056]
[0057] Among them, i is the current step number, s (i-1) is the estimated vector of the previous step, and s (0) =x n-1 , ρ is the momentum coefficient, Δs is the momentum, is the momentum update vector;
[0058] In the momentum update vector Calculate the gradient at the point to get the estimated vector of the current step, so that the estimated update is the superposition of the update direction of the previous step and the current gradient, which can be expressed as:
[0059]
[0060] Among them, s (i) is the estimated vector of the current step, (·) H is the conjugate transpose of the matrix, for The gradient at The step size of the first-order acceleration method is trainable and is obtained by training a deeply expanded Markov chain Monte Carlo detection network. The benefits include: avoiding the complex line search process of optimizing the step size in real time during the deployment phase, and more effectively approaching the global optimal solution compared to the fixed step size or the step size obtained by line search;
[0061] Calculate the increment of the current step estimate vector relative to the previous step estimate vector, that is, momentum, which can be expressed as:
[0062] Δs=s (i) -s (i1) ;
[0063] Let step number i = i + 1; if step number i > N g , then let the proposal vector mean Otherwise, project the elements of each dimension of the current step estimation vector to the transmission constellation point with the closest Euclidean distance to it, and obtain additional samples x ex , used to supplement the sample list;
[0064] Specifically, when using the second-order Newton method, the gradient descent process includes:
[0065] Calculate the objective function, that is, the Hessian matrix H of the estimated residual norm H H, get the gradient preprocessing operator Among them, γ is the guarantee of M p The coefficient of positive definiteness is taken as γ = σ in this embodiment. 2 / d 2 , d is half of the minimum Euclidean distance between transmission constellation points, YesN t ×N t dimensional identity matrix, (·) -1 is the matrix inversion operation;
[0066] Calculate the mean of the proposal vector, which can be expressed as:
[0067]
[0068] Among them, τ 2,n is the step size of the second-order Newton method, which is trainable and is obtained by training a deeply expanded Markov chain Monte Carlo detection network. The benefits it brings are similar to those of using the first-order acceleration method, which will not be repeated here.
[0069] S202, adding disturbance noise to determine the proposal vector covariance, which can be expressed as:
[0070] s p =s n +a n M c w n
[0071] Among them, s p is the proposal vector after adding perturbation noise, w n obey ), is a complex Gaussian distribution, YesN t dimensional zero vector, matrix M c By changing the matrix (H H H) -1 Perform Cholesky decomposition to obtain the covariance of the proposal vector (H H H) -1 , which is the inverse of the Hessian matrix, a n is the amplitude of the disturbance noise, which is trainable and obtained by training a deeply expanded Markov chain Monte Carlo detection network. The benefits include: overcoming the problem of poor robustness of heuristic disturbance noise amplitude setting.
[0072] S203, propose vector s p The elements of each dimension are projected to the transmission constellation point with the closest Euclidean distance to obtain the candidate sample x c , the projection uses quantization functions, acting on the real and imaginary parts of the complex-valued elements separately, including:
[0073] The approximate quantization function Γ is used in the training (such as step S3), which can be expressed as:
[0074]
[0075] Wherein, x is a real continuous value parameter of the function input, M is the cardinality of the real value transmission constellation, m is the sequence number of the constellation symbol, tanh[·] is a differentiable hyperbolic function, η is a preset fitting coefficient, which is taken as 100 in this embodiment to approximate a discontinuous step function, and E s is the average energy of the transmission constellation; using the approximate quantization function in training can avoid the problem that the gradient cannot be calculated during the error back propagation process;
[0076] In the deployment phase (such as step S4), the precise quantization function Λ is used, which can be expressed as:
[0077]
[0078] Here, ε(·) is a step function.
[0079] S204, the ratio of the likelihood probability of the candidate sample to the likelihood probability of the output sample of the previous layer of sub-sampling network is used as the acceptance probability P acc In this embodiment, the specific calculation is based on the Metropolis-Hastings (MH) criterion, which can be expressed as:
[0080]
[0081] Among them, min means taking the minimum value, l c =||r c || is the residual norm of the candidate sample, r c =y-Hx c is the residual of the candidate sample; take a random number in, Indicates uniform distribution; judge P acc Greater than or equal to P uni Is it true? If not, let the sample of the current layer sub-sampling network be the sample output by the previous layer sub-sampling network, that is, x n =x n-1 ; If so, let the sample of the current layer subsampling network be the candidate sample, that is, x n =xc ; Get the output sample x n , calculate the output sample residual r n and the sample residual norm l n .
[0082] Furthermore, in this embodiment, the training data set collected in step S1 includes 6,400,000 pieces of training data, which are divided into 5,000 training rounds, and each training round includes 10 small batches.
[0083] Furthermore, the method for training a deeply expanded Markov chain Monte Carlo detection network specifically includes:
[0084] With channel matrix H and noise variance σ 2 and the received signal y as the input features of the network, and after each layer of sub-sampling network, the steps S201-S204 are completed, and the samples of the elements of each dimension projected to the transmission constellation points are output to form a sample list
[0085] In this embodiment, the Adam optimizer is used with an initial learning rate of 0.001 to optimize the trainable parameters in the network in an unsupervised learning manner, including the step size τ of the gradient descent and the amplitude a of the perturbation noise. There is no need to use the real transmitted symbol vector x as a label. The training loss function is is the residual norm, expressed as:
[0086]
[0087] in, Represents a small batch of received signal y, channel matrix H and noise variance σ for the training data set 2 Find the mathematical expectation, is the sample in the sample list that minimizes the residual norm, which can be expressed as:
[0088]
[0089] Furthermore, the parallel sampling method described in step S4 is specifically: using multiple Markov chain Monte Carlo detection networks with the same structure and parameters and with the same depth to generate samples for the same channel to form a sample list. In this embodiment, a hard decision is made based on the sample list, and the sample with the smallest residual norm is selected as the estimation result of the transmitted symbol vector.
[0090] 3. Implementation Effect
[0091] The Markov chain Monte Carlo detection method in the embodiment of the present invention is compared with similar technical methods. The results are as follows: Figure 3 and Figure 4 As shown:
[0092] Figure 3 The BER comparison effect of the present invention and the linear minimum mean square error (MMSE) detection, maximum likelihood detection and the prototype algorithm of the present invention when the gradient descent part uses the first-order acceleration method is given. Among them, the prototype algorithm of the present invention is a gradient descent step size of The disturbance noise amplitude is taken as Markov chain Monte Carlo detection network, where ||·|| F Represents the Frobenius norm of the matrix, and max represents the maximum value. In the legend, P is the number of parallel detection networks used, and N is the number of sub-sampling networks contained in the detection network. It can be seen that under different N values, the performance of the present invention is improved compared with the prototype algorithm, and the problem of the upward BER curve of the prototype algorithm in the high signal-to-noise ratio (SNR) area is effectively overcome. At the same time, compared with MMSE detection, the present invention has a significant gain and can achieve performance close to maximum likelihood detection with a lower computational overhead.
[0093] Figure 4 The BER comparison effect of the present invention and MMSE detection, maximum likelihood detection and the prototype algorithm of the present invention when the gradient descent part uses the second-order Newton method is given. Among them, the prototype algorithm of the present invention is a gradient descent step size of The disturbance noise amplitude is taken as Markov chain Monte Carlo detection network, where Re(·) represents the real part, and the matrix G = HM p H H . The initial sample x0 of the first layer subsampling network input of the present invention and the prototype algorithm adopts different initialization methods. Among them, the present invention uses a low-complexity random initialization method, and the prototype algorithm uses random initialization and high-complexity MMSE initialization methods, which are marked with "-Random" and "-MMSE" respectively in the legend. It can be seen that under different values of N, the present invention has improved performance compared with the randomly initialized prototype algorithm. With a simple initialization method, it can achieve performance comparable to or even better than that of the MMSE initialized prototype algorithm, while avoiding the complex line search process of calculating the gradient descent step length, and has the advantage of low computational overhead.
[0094] What is described above in conjunction with the accompanying drawings is only a preferred embodiment of the present invention and cannot be used to limit the scope of rights included in the present invention. It should be understood that any equivalent changes made without departing from the spirit of the present invention are within the scope of protection covered by the claims of the present invention.
Claims
1. A Markov chain Monte Carlo detection method based on deep expansion, characterized in that: The method comprises the following steps: S1. Collect a training data set of the MIMO wireless communication system, where each piece of training data includes a channel matrix and noise variance of a channel realization, and a corresponding received signal vector; S2, construct a deeply expanded Markov chain Monte Carlo detection network; S3, using the collected training data set to train a deeply expanded Markov chain Monte Carlo detection network; S4. In the deployment phase, the trained deep expanded Markov chain Monte Carlo detection network is used to obtain a sample list by parallel sampling. A judgment is made based on the sample list to obtain a sample with the smallest residual norm in the sample list or a log-likelihood ratio as a detection result. The deeply expanded Markov chain Monte Carlo detection network comprises N layers of sub-sampling networks with the same structure, wherein the input of each layer of sub-sampling network comprises: a channel matrix, a noise variance and a received signal vector, as well as samples, sample residuals and sample residual norms output by the previous layer of sub-sampling network; the output of each layer of sub-sampling network comprises: samples, sample residuals and sample residual norms of the current layer; Each layer of the sub-sampling network of the deeply expanded Markov chain Monte Carlo detection network specifically includes the following steps: S201, executing a gradient descent process to calculate a mean of a proposal vector, wherein the step size of the gradient descent is trainable and is obtained by training a deeply expanded Markov chain Monte Carlo detection network; S202, adding disturbance noise to determine the covariance of the proposal vector, wherein the amplitude of the disturbance noise is trainable and is obtained by training a deeply expanded Markov chain Monte Carlo detection network; S203, projecting the elements of each dimension of the proposal vector to the transmission constellation point with the closest Euclidean distance to it, to obtain candidate samples; S204, the ratio of the likelihood probability of the candidate sample to the likelihood probability of the output sample of the previous layer of sub-sampling network is used as the acceptance probability P acc ; Get a random number in, Indicates uniform distribution; judge P acc Greater than or equal to P uni Is it true? If not, let the samples of the current layer sub-sampling network be the samples output by the previous layer sub-sampling network; if yes, let the samples of the current layer sub-sampling network be the candidate samples and get the output samples; calculate the output sample residual and sample residual norm; The training method of the deeply expanded Markov chain Monte Carlo detection network specifically includes: The channel matrix, noise variance and received signal are used as input features of the network, and the steps S201-S204 are completed through each layer of sub-sampling network, and samples of the elements of each dimension projected to the transmission constellation points are output to form a sample list; Optimize the trainable parameters in the network in an unsupervised learning manner, including the step size of gradient descent and the amplitude of perturbation noise, and the loss function of training is the residual norm, expressed as: Where y represents the received signal, H represents the channel matrix, σ 2 represents the noise variance, represents the mathematical expectation of a small batch of received signals, channel matrices, and noise variances for the training data set, ||·|| represents the norm of the vector, is the sample in the sample list that minimizes the residual norm; The method of projecting to the transmission constellation point uses a quantization function to act on the real part and the imaginary part of the complex value element respectively, specifically including: The approximate quantization function Γ is used in training, and its expression is: Where x is the real continuous value parameter of the function input, M is the cardinality of the real value transmission constellation, m is the sequence number of the constellation symbol, tanh[·] is the hyperbolic function, η is the preset fitting coefficient, and E s is the average energy of the transmission constellation; In the deployment phase, the precise quantization function Λ is used, which is expressed as: Here, ε(·) is a step function.
2. The Markov chain Monte Carlo detection method based on deep expansion according to claim 1 is characterized in that: The gradient descent process described in step S201 uses a first-order acceleration method or a second-order Newton method; when using the first-order acceleration method, the gradient descent process includes multiple steps, and at the end of each step, the elements of each dimension of the current step estimation vector are projected to the transmission constellation point with the closest Euclidean distance to it to obtain additional samples.
3. The Markov chain Monte Carlo detection method based on deep expansion according to claim 1 is characterized in that: The parallel sampling method described in step S4 is specifically: using multiple Markov chain Monte Carlo detection networks with the same structure and parameters and with the same depth to generate samples for the same channel to form a sample list.
Citation Information
Patent Citations
Reduced complexity detection and decoding for a receiver in a communication system
CN101322328A
Approximate message passing large-scale MIMO signal detection algorithm based on deep neural network
CN111224906A
