Mixing precision quantization method and device for MIMO (Multiple Input Multiple Output) detection

By initializing the bit width of the integer and decimal parts of the variables to be quantized in MIMO detection, and using the agent to automatically optimize, the problem of insufficient compression of the bit width of the integer variable is solved, and hardware implementation with lower complexity and more efficient resource utilization is achieved.

CN119675787APending Publication Date: 2025-03-21PURPLE MOUNTAIN LAB
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411716801.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-27
Publication Date
2025-03-21

AI Technical Summary

Technical Problem

The prior art lacks compression of integer variable bit widths in MIMO detection, and cannot find the optimal quantized bit width, resulting in high complexity in hardware implementation.

Method used

By obtaining the variables to be quantized in the MIMO signal detection process, mapping them into integers and decimal parts, initializing the bit width based on the probability density function and unified quantization method, and using the agent of the deep reinforcement learning algorithm for training, the target quantization bit width is automatically optimized.

Benefits of technology

While ensuring detection performance, it further reduces the complexity of hardware implementation, improves the bit width compression rate, reduces hardware resource consumption, and improves area efficiency and energy efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119675787A_ABST
    Figure CN119675787A_ABST
Patent Text Reader

Abstract

The invention provides a mixed precision quantization method and device for MIMO detection, and belongs to the technical field of communication, and the method comprises the steps: mapping a variable to be quantized in an MIMO signal detection process into an integer variable to be quantized and a decimal variable to be quantized; determining an integer initialization bit width corresponding to each integer to-be-quantized variable based on the probability density function, and determining a decimal initialization bit width corresponding to each decimal to-be-quantized variable based on a unified quantization method; training the intelligent agent according to the integer initialization bit width and the decimal initialization bit width, and determining a target quantization bit width output by the intelligent agent; and according to the target quantization bit width, simultaneously quantizing the integer to-be-quantized variable and the decimal to-be-quantized variable to obtain a quantization result. According to the method, the to-be-quantized variables are mapped into two parts, the integers in the DRL algorithm are used for automatically optimizing the integer to-be-quantized variables and the decimal to-be-quantized variables in the training process at the same time, the bit width can be further reduced, and the complexity of hardware implementation is further reduced while the detection performance is guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a mixed precision quantization method and device for MIMO detection. Background Art

[0002] With the rapid development of 5G wireless communication technology, multiple-input multiple-output (MIMO) plays an important role in improving channel capacity, coverage, transmission rate, etc. Due to the high complexity of large-scale MIMO, in order to reduce the demand for computing and storage and facilitate hardware implementation and deployment, MIMO detection quantization is usually used to eliminate redundant bit width while ensuring bit error rate (BER) performance.

[0003] The existing automatic hybrid-precision quantization (AHPQ) algorithm for MIMO detection quantization uses deep reinforcement learning (DRL) to automatically optimize the fractional part (FP) of the variable. For the integer part (IP), there is still room for bit width compression, and the optimal quantization bit width cannot be found, resulting in high complexity in hardware implementation. Summary of the invention

[0004] The present invention provides a mixed precision quantization method and device for MIMO detection, which are used to solve the defects of the prior art that the bit width of integer variables is insufficiently compressed, the optimal quantization bit width cannot be found, and the hardware implementation has high complexity.

[0005] The present invention provides a mixed precision quantization method for MIMO detection, comprising: Obtaining variables to be quantized in a multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized; Determine the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determine the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; The agent is trained according to the integer initialization bit width and the decimal initialization bit width, and a target quantization bit width output by the agent corresponding to the integer variable to be quantized and the decimal variable to be quantized is determined; The integer variable to be quantized and the fractional variable to be quantized are quantized simultaneously according to the target quantization bit width to obtain a quantization result.

[0006] According to a mixed precision quantization method for MIMO detection provided by the present invention, the training of the agent according to the integer initialization bit width and the decimal initialization bit width includes: Determine the state, action and reward function of the agent according to the integer initialization bit width and the decimal initialization bit width; The agent is trained according to the state, action and reward function of the agent.

[0007] According to a mixed precision quantization method for MIMO detection provided by the present invention, the state, action and reward function of the agent are determined according to the integer initialization bit width and the decimal initialization bit width, including: Extracting a target variable from the integer variables to be quantified and the decimal variables to be quantified, and determining the state of the agent according to the target variable; Determine the action of the agent according to the maximum value of each of the integer initialization bit widths and the decimal initialization bit width, the initialization bit width corresponding to the target variable, and a predetermined bit width change value generated after the agent takes an action; The reward function of the agent is determined according to the optimization goal of the agent and the preset hyperparameters.

[0008] According to a mixed precision quantization method for MIMO detection provided by the present invention, the intelligent agent is trained according to the state, action and reward function of the intelligent agent, including: During a current round of training of the agent, randomly selecting a variable from the pre-extracted target variables; The agent interacts with the environment, takes actions corresponding to the extracted variables, determines a reward value corresponding to the extracted variables according to the reward function, and updates the state; If the reward value meets the preset conditions, enter the next round of training; if the reward value does not meet the preset conditions, update the target quantization bit width corresponding to the extracted variable and enter the next round of training until the target quantization bit width corresponding to the integer variable to be quantized and the decimal variable to be quantized are determined.

[0009] According to a mixed precision quantization method for MIMO detection provided by the present invention, in the training process of the intelligent agent, the method further includes: During the current round of training of the agent, performing an action according to a current state and a current variable to obtain a current variable bit width corresponding to the current variable; If the bit width of the current variable meets the preset termination state, stop the current round of training and update the current variable; The termination state is determined according to the correlation between the variables to be quantized, and the correlation between the variables to be quantized is used to characterize the influence of the bit width of a certain variable to be quantized on the bit width of other variables to be quantized.

[0010] According to a mixed precision quantization method for MIMO detection provided by the present invention, the integer variables to be quantized correspond to at least one integer bit width value, and correspondingly, determining the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determining the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method includes: According to the value range of the probability density function corresponding to the different integer variables to be quantized and the integer bit width value, the constraints of the different integer variables to be quantized are determined; the minimum value of the integer bit width value that satisfies the constraints of the integer variables to be quantized is used as the integer initialization bit width corresponding to the integer variables to be quantized; the same decimal bit width value is set for the different decimal variables to be quantized, the decimal bit width value is adjusted and the relative bit error rate before and after the test quantization corresponding to the adjusted decimal bit width value is determined, and the minimum bit width value corresponding to the corresponding bit error rate less than a preset threshold is selected as the decimal initialization bit width. The present invention also provides a mixed precision quantization device for MIMO detection, comprising: A mapping module, used for obtaining variables to be quantized in a multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized; A first determination module is used to determine the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and to determine the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; A second determination module is used to train the agent according to the integer initialization bit width and the decimal initialization bit width, and determine the target quantization bit width output by the agent corresponding to the integer variable to be quantized and the decimal variable to be quantized respectively; The quantization module is used to simultaneously quantize the integer variable to be quantized and the decimal variable to be quantized according to the target quantization bit width to obtain a quantization result.

[0011] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the mixed precision quantization method for MIMO detection as described in any one of the above is implemented.

[0012] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the mixed precision quantization method for MIMO detection as described in any one of the above is implemented.

[0013] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the mixed precision quantization method for MIMO detection as described in any one of the above is implemented.

[0014] The mixed precision quantization method and device for MIMO detection provided by the present invention obtain variables to be quantized in the process of multi-input multi-output MIMO signal detection, and map each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized respectively; determine the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determine the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; train an intelligent agent according to the integer initialization bit width and the decimal initialization bit width, and determine the target quantization bit width output by the intelligent agent corresponding to each of the integer variable to be quantized and the decimal variable to be quantized; and simultaneously quantize the integer variable to be quantized and the decimal variable to be quantized according to the target quantization bit width to obtain a quantization result. The present invention maps the variables to be quantized into two parts, initializes the bit width of each integer variable to be quantized through a probability density function, initializes the bit width of each decimal variable to be quantized based on a unified quantization method, and uses an intelligent agent in a DRL algorithm to automatically optimize the integer variables to be quantized and the decimal variables to be quantized during training, so that the bit width of the integer variables to be quantized can be further reduced on the original basis, so as to further reduce the complexity of hardware implementation while ensuring the detection performance, and solves the defects of insufficient compression of the bit width of integer variables, inability to find the optimal quantization bit width, and high complexity of hardware implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a flow chart of the mixed precision quantization method for MIMO detection provided by the present invention.

[0017] Figure 2 It is a schematic diagram of the mapping process of the variables to be quantified provided by the present invention.

[0018] Figure 3 It is a structural schematic diagram of the MMSE detector based on LU decomposition provided by the present invention.

[0019] Figure 4 It is a schematic diagram of the training process of the intelligent agent provided by the present invention.

[0020] Figure 5 It is a schematic diagram of the introduction termination state provided by the present invention.

[0021] Figure 6 It is a performance simulation schematic diagram provided by the present invention.

[0022] Figure 7 The figure is a schematic diagram of the structure of a mixed precision quantization device for MIMO detection provided by the present invention.

[0023] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] As one of the commonly used MIMO detection quantization methods, unified quantization (UQ) directly allocates bit widths to each variable based on predefined system performance requirements. However, such a rough quantization method will cause considerable bit redundancy; another commonly used MIMO detection quantization method is manual selection (MS), which allocates a specific bit width to each variable through the designer's expert experience. Although this method can reduce bit redundancy to a certain extent, it takes a lot of time and manpower and material resources; the third automatic mixed-precision quantization (AHPQ) algorithm uses deep reinforcement learning (DRL) to automatically optimize the fractional part (FP) of the variable. Therefore, there is still room for bit width compression for the integral part (IP). In addition, this method lacks consideration of the correlation between variables, which may lead to slow and unstable DRL training and failure to find the optimal quantization bit width.

[0026] Based on the DRL algorithm, the present invention proposes a method called generalized automatic hybrid precision quantization (GausiQ), which can efficiently and automatically explore the optimal quantization scheme, further compress the bit width while ensuring performance, reduce hardware resource consumption, improve area efficiency and energy efficiency, and further reduce the complexity of hardware implementation while ensuring detection performance. The present invention is described in detail below in conjunction with the accompanying drawings.

[0027] Figure 1 is a flow chart of the mixed precision quantization method for MIMO detection provided by the present invention, such as Figure 1 As shown, the present invention provides a mixed precision quantization method for MIMO detection, comprising the following steps.

[0028] Step S100, obtaining variables to be quantized in the process of multi-input multi-output MIMO signal detection, and mapping each of the variables to be quantized into an integer variable to be quantized and a fractional variable to be quantized. The variables to be quantized refer to the variables to be quantized generated by the MIMO signal detector in the process of signal detection of the MIMO system, including input variables, intermediate variables and output variables of the MIMO signal detector, and each variable to be quantized includes an integer variable to be quantized IP and a fractional integer variable to be quantized FP based on the determination of the MIMO signal detector.

[0029] Step S200, determining the integer initialization bit width corresponding to each of the integer variables to be quantized based on the probability density function and determining the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on the unified quantization method. The present invention initializes the bit width of the integer variables to be quantized based on the probability density function (PDF) to obtain the integer initialization bit width; and initializes the decimal variables to be quantized based on the unified quantization method UQ to obtain the decimal initialization bit width. It should be noted that the decimal initialization bit width corresponding to each of the decimal variables to be quantized determined based on the unified quantization method is the same bit width value.

[0030] Step S300, training an agent according to the integer initialization bit width and the decimal initialization bit width, and determining a target quantization bit width output by the agent corresponding to each of the integer variable to be quantized and the decimal variable to be quantized.

[0031] The present invention trains an intelligent agent based on a DRL algorithm, and interaction data (such as state, action value, and reward value) during the intelligent agent training process is determined based on the integer initialization bit width and the decimal initialization bit width.

[0032] Step S400 , quantizing the integer variable to be quantized and the fractional variable to be quantized simultaneously according to the target quantization bit width to obtain a quantization result.

[0033] Simultaneously quantizing the integer variable to be quantized and the decimal variable to be quantized according to the target quantization bit width specifically refers to quantizing the integer variable to be quantized or the decimal variable to be quantized according to the target quantization bit widths corresponding to each of the integer variable to be quantized and the decimal variable to be quantized, and using the quantized integer variable to be quantized and the decimal variable to be quantized as quantization results.

[0034] It can be understood that the present invention maps the variables to be quantized into two parts, initializes the bit width of each integer variable to be quantized through a probability density function, initializes the bit width of each decimal variable to be quantized based on a unified quantization method, and utilizes the intelligent agent in the DRL algorithm to automatically optimize the integer variables to be quantized and the decimal variables to be quantized during the training process, thereby solving the problem of insufficient compression of the bit width of integer variables and inability to find the optimal quantization bit width.

[0035] Figure 2 is a schematic diagram of the mapping process of the variables to be quantified provided by the present invention, such as Figure 2 As shown, based on the above embodiment, as an optional embodiment, the integer variables to be quantized correspond to at least one integer bit width value, and correspondingly, the integer initialization bit width corresponding to each of the integer variables to be quantized is determined based on the probability density function and the decimal initialization bit width corresponding to each of the decimal variables to be quantized is determined based on the unified quantization method, including the following steps.

[0036] According to the value range of the probability density function corresponding to different integer variables to be quantized and the integer bit width value, the constraints of different integer variables to be quantized are determined; the minimum value of the integer bit width value that satisfies the constraints of the integer variables to be quantized is used as the integer initialization bit width corresponding to the integer variables to be quantized; the same decimal bit width value is set for different decimal variables to be quantized, the decimal bit width value is adjusted and the relative bit error rate before and after the test quantization corresponding to the adjusted decimal bit width value is determined, and the minimum bit width value corresponding to the corresponding bit error rate less than a preset threshold is selected as the decimal initialization bit width.

[0037] Assumptions The variables to be quantified constitute a set , to collect A variable in For example, the variable Mapped to IP variables to be quantified FP variables to be quantized , forming a set , the variable and The bit widths are recorded as and By By quantizing the variables in , you can achieve simultaneous quantization of the variable IP and FP parts.

[0038] The calculation formula for determining the integer initialization bit width corresponding to each of the integer variables to be quantized based on the value range of the probability density function corresponding to the different integer variables to be quantized is as follows: ; in, Initialize the bit width for integers, are multiple sets of samples generated by Monte Carlo simulation, for Falling in the interval The multiset of elements outside The operation counts the number of elements in a multi-set. is a predefined threshold.

[0039] The calculation formula for determining the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on the unified quantization method is as follows: ; in, Initialize the bit width for the decimal, To quantify the former BER loss, is the BER loss after quantization, is the relative BER loss, v It is a relative bit error rate threshold set manually.

[0040] Figure 3 is a schematic diagram of the structure of the MMSE detector based on LU decomposition provided by the present invention, such as Figure 3 As shown, taking the MMSE detector based on LU decomposition commonly used in the MIMO signal detection process as an example, the channel matrix , received signal vector , noise matrix , Graham Matrix , the received signal vector after matched filtering , intermediate variable matrix ,Depend on Decomposed lower triangular matrix With the upper triangular matrix , intermediate variables , the estimated transmitted signal vector These are variables involved in the MMSE detector using LU decomposition, and are also variables to be quantized in the quantization operation of the present invention. The index of the variable to be quantized and its corresponding quantizer are shown in the following table.

[0041] The elements in the matrix or vector corresponding to the variables in the table use the same quantization scheme. To facilitate hardware implementation, the variables corresponding to the seventh quantizer use the same quantization scheme.

[0042] Taking the MMSE detector using LU decomposition as an example, a large number of samples are generated based on Monte Carlo simulation, and these samples can fit the probability density function of the variables in the above table. According to the different value ranges of the probability density functions of different variables, different initialization integer bit widths of different variables can be determined. Set the decimal bit widths of different variables to the same value, uniformly adjust the decimal bit widths, test the relative bit error rate before and after quantization, and select the minimum uniform bit width that can meet the relative bit error rate less than the threshold as the decimal initialization bit width according to the artificially set relative bit error rate threshold.

[0043] It can be understood that the present invention uses a probability density function to determine the constraints for initializing the bit width of integer variables to be quantized, and initializes the bit width of all decimal variables to be quantized through a unified quantization method. The decimal initialization bit width of all decimal variables to be quantized is the same, and the variables to be quantized are mapped into two independent parts. The integer variables to be quantized and the decimal variables to be quantized are automatically optimized, which solves the problem of insufficient compression of the bit width of integer variables and the inability to find the optimal quantization bit width.

[0044] Based on the above embodiment, as an optional embodiment, the training of the intelligent agent according to the integer initialization bit width and the decimal initialization bit width includes the following steps.

[0045] Step S310, determining the state, action and reward function of the agent according to the integer initialization bit width and the decimal initialization bit width. The reward function of the agent is related to the optimization target of the agent.

[0046] Step S320, training the agent according to the state, action and reward function of the agent.

[0047] It can be understood that the present invention regards the IP and FP of the variable to be quantized as two independent parts, and substitutes the initialization bit widths corresponding to the IP and FP of the variable to be quantized into the training process of the intelligent agent, which is conducive to the intelligent agent to simultaneously optimize the IP and FP of the variable to be quantized and improve the bit width compression rate.

[0048] Based on the above embodiment, as an optional embodiment, determining the state, action and reward function of the agent according to the integer initialization bit width and the decimal initialization bit width includes the following steps.

[0049] Step S311, extracting a target variable from the integer variables to be quantified and the decimal variables to be quantified, and determining the state of the agent according to the target variable.

[0050] Assume that from the set Extract the target variable , , is the number of target variables, the state can be defined as , The operation represents one-hot encoding. The target variable The variable type is not limited. It is possible to extract integer variables to be quantified or decimal variables to be quantified.

[0051] Step S312, determining the action of the agent according to the maximum value of each of the integer initialization bit widths and the decimal initialization bit widths, the initialization bit width corresponding to the target variable, and a predetermined bit width change value generated after the agent takes an action.

[0052] The bit-width expression after the agent takes an action is as follows: ; in, is the bit width after the agent takes action, The target variable The corresponding initialization bit width, is the bit width change value generated after the agent takes action, is the maximum value of the integer initialization bit width and the decimal initialization bit width. The action space of the agent is ,in, is the maximum value of the bit width change.

[0053] Step S313, determining the reward function of the agent according to the optimization goal of the agent and the preset hyperparameters.

[0054] The expression of the reward function is as follows: in, Indicates the average bit width of the variables extracted to reduce the correlation between variables, which is selected from the optimization goal. The average value of the bit width of the target variables is obtained based on the optimization goal of the agent. and is a hyperparameter.

[0055] The optimization goal of the intelligent agent is defined as: At the same time, the average bit width is minimized. The expression of the agent's optimization objective is as follows: ; in, is the optimization goal of the agent, for Middle j The intermediate bit width of a variable during the agent training process.

[0056] It can be understood that the present invention provides a specific solution for determining the state, action and reward function of the agent, extracting the target variable from all variables, reducing the correlation between variables, and improving the accuracy of the agent.

[0057] Based on the above embodiment, as an optional embodiment, the training of the agent according to the state, action and reward function of the agent includes the following steps.

[0058] Step S321, during the current round of training of the agent, a variable is randomly selected from the pre-extracted target variables.

[0059] Optionally, the number of iterations may be directly set to allow the agent to directly perform iterative training, or nested iteration conditions may be set to allow the agent to perform iterative training when the nested iteration conditions are met.

[0060] Figure 4 It is a schematic diagram of the training process of the intelligent agent provided by the present invention, such as Figure 4 As shown, the nested iteration condition includes the first judgment parameter , the second judgment parameter and the third decision parameter , the first judgment parameter Number of updates by policy and policy update interval The second decision parameter is determined by Update interval by policy OK, the third judgment parameter By time step If the first judgment parameter , then reduce its value by 1 and use it as the second judgment parameter Assign value, initialize the bit width counter HisVar , if the second judgment parameter , then reduce its value by 1 and use it as the third judgment parameter Assignment, initialize the environment, extract the target variable from the collection, if the third judgment parameter , then reduce its value by 1 and start the current round of training for the agent.

[0061] Step S322, the agent interacts with the environment, takes actions corresponding to the extracted variables, determines the reward value corresponding to the extracted variables according to the reward function, and updates the state.

[0062] Step S323: If the reward value meets the preset conditions, enter the next round of training; if the reward value does not meet the preset conditions, update the target quantization bit width corresponding to the extracted variable and enter the next round of training until the target quantization bit width corresponding to the integer variable to be quantized and the decimal variable to be quantized are determined.

[0063] If the reward value does not meet the preset conditions (such as Figure 3 The reward value in is not equal to -1), update the target quantization bit width corresponding to the extracted variable and enter the next round of training.

[0064] Through the number of iterations or nested iteration conditions, the agent training can be obtained The target quantization bit width of all variables in is the bit width counter HisVar The bit width value with the largest number of occurrences corresponding to each variable in .

[0065] It can be understood that the present invention provides a specific solution for training intelligent agents. By determining the target quantization bit width of each variable, the bit width is greatly compressed, the hardware resource consumption is reduced, the area efficiency and energy efficiency are improved, and the defect of insufficient compression of integer bit width in the prior art is solved.

[0066] Based on the above embodiment, as an optional embodiment, the training process of the intelligent agent further includes the following steps.

[0067] During the current round of training of the agent, actions are performed according to the current state and the current variables to obtain a current variable bit width corresponding to the current variables.

[0068] If the bit width of the current variable meets the preset termination state, the current round of training is stopped and the current variable is updated.

[0069] The termination state is determined according to the correlation between the variables to be quantized, and the correlation between the variables to be quantized is used to characterize the influence of the bit width of a certain variable to be quantized on the bit width of other variables to be quantized.

[0070] In the embodiment of the present invention, the termination state refers to that the reward value corresponding to the current variable bit width is -1. Correspondingly, if the reward value corresponding to the current variable bit width is -1, it is determined that the current variable bit width meets the termination state, and the current round of training is stopped.

[0071] Figure 5 It is a schematic diagram of the introduction termination state provided by the present invention, such as Figure 5 As shown, unlike the common neural network that introduces a termination state / early stopping, the present invention innovatively introduces a termination state that takes into account the influence of the correlation between variables.

[0072] The so-called correlation between variables means that the quantization bit width of one variable may affect the quantization of other variables. Figure 5 For example, after resetting, the agent will Select the second variable and take action After taking action, the bit width of the second variable is updated to , which is lower than the ideal bit width. Therefore, the environment returns a reward value . Assume that the status is updated to , then no matter how the agent takes action to change the bit width of the first variable (for example Figure 2 take ), the environment will return The reward value.

[0073] By introducing the terminal state, we can avoid the agent from In other words, the environment will be reset and a new round of training will start. When the variable bit width meets the termination state, the current round of training will be stopped and restarted. Extract The variables are quantified to avoid misjudgment of reward values, fully consider different variable subsets, and promote the exploration of the optimization process.

[0074] It can be understood that in order to further consider the correlation between variables, the present invention introduces a termination state. The introduction of the termination state can fully avoid the impact of the correlation between variables, promote the intelligent agent's exploration of different quantization schemes during the optimization process, and improve the sampling efficiency.

[0075] The performance of the present invention is illustrated below in conjunction with preferred embodiments.

[0076] Application of the present invention to a transmitting antenna , receiving antenna , a massive MIMO system with 16-QAM modulation and minimum mean square error (MMSE) detection is used. The default value is 0.5. Test BER performance at the

[0077] Figure 6 It is a performance simulation schematic diagram provided by the present invention, such as Figure 6As shown, the horizontal axis is BER performance, and the vertical axis is average quantization bit width (which can be understood as the average value of all target quantization bit widths). The present invention can reduce the average quantization bit width to , ; Compared with the technology without introducing the termination state, the final quantization result and the number of extracted variables The value of is independent and has a smaller relative SNR loss, avoiding cumbersome and complex The selection of reduces the amount of calculation.

[0078] The mixed precision quantization device for MIMO detection provided by the present invention is described below. The mixed precision quantization device for MIMO detection described below and the mixed precision quantization method for MIMO detection described above can refer to each other.

[0079] Figure 7 is a schematic diagram of the structure of a mixed precision quantization device for MIMO detection provided by the present invention, such as Figure 7 As shown, the present invention also provides a mixed precision quantization device for MIMO detection, including a sender and a receiver, and the sender includes the following modules.

[0080] A mapping module 710 is used to obtain variables to be quantized in a multi-input multi-output MIMO signal detection process, and map each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized; A first determining module 720 is used to determine the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determine the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; A second determination module 730 is used to train the agent according to the integer initialization bit width and the decimal initialization bit width, and determine the target quantization bit width output by the agent corresponding to the integer variable to be quantized and the decimal variable to be quantized respectively; The quantization module 740 is used to simultaneously quantize the integer variable to be quantized and the fractional variable to be quantized according to the target quantization bit width to obtain a quantization result.

[0081] As an embodiment, the second determining module 730 is further configured to: Determine the state, action and reward function of the agent according to the integer initialization bit width and the decimal initialization bit width; The agent is trained according to the state, action and reward function of the agent.

[0082] As an embodiment, the second determining module 730 is further configured to: Extracting a target variable from the integer variables to be quantified and the decimal variables to be quantified, and determining the state of the agent according to the target variable; Determine the action of the agent according to the maximum value of each of the integer initialization bit widths and the decimal initialization bit width, the initialization bit width corresponding to the target variable, and a predetermined bit width change value generated after the agent takes an action; The reward function of the agent is determined according to the optimization goal of the agent and the preset hyperparameters.

[0083] As an embodiment, the second determining module 730 is further configured to: During a current round of training of the agent, randomly selecting a variable from the pre-extracted target variables; The agent interacts with the environment, takes actions corresponding to the extracted variables, determines a reward value corresponding to the extracted variables according to the reward function, and updates the state; If the reward value meets the preset conditions, enter the next round of training; if the reward value does not meet the preset conditions, update the target quantization bit width corresponding to the extracted variable and enter the next round of training until the target quantization bit width corresponding to the integer variable to be quantized and the decimal variable to be quantized are determined.

[0084] As an embodiment, the second determining module 730 is further configured to: During the current round of training of the agent, performing an action according to a current state and a current variable to obtain a current variable bit width corresponding to the current variable; If the bit width of the current variable meets the preset termination state, stop the current round of training and update the current variable; The termination state is determined according to the correlation between the variables to be quantized, and the correlation between the variables to be quantized is used to characterize the influence of the bit width of a certain variable to be quantized on the bit width of other variables to be quantized.

[0085] As an embodiment, the integer variable to be quantized corresponds to at least one integer bit width value, and correspondingly, the second determining module 730 is further used to: According to the value range of the probability density function corresponding to different integer variables to be quantized and the integer bit width value, the constraints of different integer variables to be quantized are determined; the minimum value of the integer bit width value that satisfies the constraints of the integer variables to be quantized is used as the integer initialization bit width corresponding to the integer variables to be quantized; the same decimal bit width value is set for different decimal variables to be quantized, the decimal bit width value is adjusted and the relative bit error rate before and after the test quantization corresponding to the adjusted decimal bit width value is determined, and the minimum bit width value corresponding to the corresponding bit error rate less than a preset threshold is selected as the decimal initialization bit width Figure 8 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 8 As shown, the electronic device may include: a processor (processor) 810, a communication interface (Communications Interface) 820, a memory (memory) 830 and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call the logic instructions in the memory 830 to execute the mixed precision quantization method for MIMO detection, the method including: obtaining the variables to be quantized in the multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized respectively; determining the integer initialization bit width corresponding to each of the integer variables to be quantized based on the probability density function and determining the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on the unified quantization method; training the intelligent agent according to the integer initialization bit width and the decimal initialization bit width, and determining the target quantization bit width output by the intelligent agent corresponding to the integer variable to be quantized and the decimal variable to be quantized respectively; and simultaneously quantizing the integer variable to be quantized and the decimal variable to be quantized according to the target quantization bit width to obtain a quantization result.

[0086] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0087] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the mixed precision quantization method for MIMO detection provided by the above methods, and the method includes: obtaining the variables to be quantized in the multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized respectively; determining the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determining the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; training an intelligent agent according to the integer initialization bit width and the decimal initialization bit width, and determining the target quantization bit width output by the intelligent agent corresponding to each of the integer variables to be quantized and the decimal variables to be quantized; and simultaneously quantizing the integer variables to be quantized and the decimal variables to be quantized according to the target quantization bit width to obtain a quantization result.

[0088] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the mixed precision quantization method for MIMO detection provided by the above methods, the method comprising: obtaining variables to be quantized in the multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized respectively; determining an integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determining a decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; training an intelligent agent according to the integer initialization bit width and the decimal initialization bit width, and determining a target quantization bit width output by the intelligent agent corresponding to each of the integer variables to be quantized and the decimal variables to be quantized; and simultaneously quantizing the integer variables to be quantized and the decimal variables to be quantized according to the target quantization bit width to obtain a quantization result.

[0089] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0090] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mixed precision quantization method for MIMO detection, characterized in that: include: Obtaining variables to be quantized in a multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized; Determine the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and determine the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; The agent is trained according to the integer initialization bit width and the decimal initialization bit width, and a target quantization bit width output by the agent corresponding to the integer variable to be quantized and the decimal variable to be quantized is determined; The integer variable to be quantized and the fractional variable to be quantized are quantized simultaneously according to the target quantization bit width to obtain a quantization result.

2. The mixed precision quantization method for MIMO detection according to claim 1, characterized in that: The training of the agent according to the integer initialization bit width and the decimal initialization bit width comprises: Determine the state, action and reward function of the agent according to the integer initialization bit width and the decimal initialization bit width; The agent is trained according to the state, action and reward function of the agent.

3. The mixed precision quantization method for MIMO detection according to claim 2, characterized in that: The determining, according to the integer initialization bit width and the decimal initialization bit width, the state, action and reward function of the agent comprises: Extracting a target variable from the integer variables to be quantified and the decimal variables to be quantified, and determining the state of the agent according to the target variable; Determine the action of the agent according to the maximum value of each of the integer initialization bit widths and the decimal initialization bit width, the initialization bit width corresponding to the target variable, and a predetermined bit width change value generated after the agent takes an action; The reward function of the agent is determined according to the optimization goal of the agent and the preset hyperparameters.

4. The mixed precision quantization method for MIMO detection according to claim 2, characterized in that: The step of training the agent according to the state, action and reward function of the agent comprises: During a current round of training of the agent, randomly selecting a variable from the pre-extracted target variables; The agent interacts with the environment, takes actions corresponding to the extracted variables, determines a reward value corresponding to the extracted variables according to the reward function, and updates the state; If the reward value meets the preset conditions, enter the next round of training; if the reward value does not meet the preset conditions, update the target quantization bit width corresponding to the extracted variable and enter the next round of training until the target quantization bit width corresponding to the integer variable to be quantized and the decimal variable to be quantized are determined.

5. The mixed precision quantization method for MIMO detection according to any one of claims 1 to 4, characterized in that: The training process of the intelligent agent also includes: During the current round of training of the agent, performing an action according to a current state and a current variable to obtain a current variable bit width corresponding to the current variable; If the bit width of the current variable meets the preset termination state, stop the current round of training and update the current variable; The termination state is determined according to the correlation between the variables to be quantized, and the correlation between the variables to be quantized is used to characterize the influence of the bit width of a certain variable to be quantized on the bit width of other variables to be quantized.

6. The mixed precision quantization method for MIMO detection according to claim 5, characterized in that: The integer variables to be quantized correspond to at least one integer bit width value, and correspondingly, determining the integer initialization bit width corresponding to each of the integer variables to be quantized based on the probability density function and determining the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on the unified quantization method includes: According to the value range of the probability density function corresponding to different integer variables to be quantized and the integer bit width value, the constraints of different integer variables to be quantized are determined; the minimum value of the integer bit width value that satisfies the constraints of the integer variables to be quantized is used as the integer initialization bit width corresponding to the integer variables to be quantized; the same decimal bit width value is set for different decimal variables to be quantized, the decimal bit width value is adjusted and the relative bit error rate before and after the test quantization corresponding to the adjusted decimal bit width value is determined, and the minimum bit width value corresponding to the corresponding bit error rate less than a preset threshold is selected as the decimal initialization bit width.

7. A mixed precision quantization device for MIMO detection, characterized in that: include: A mapping module, used for obtaining variables to be quantized in a multi-input multi-output MIMO signal detection process, and mapping each of the variables to be quantized into an integer variable to be quantized and a decimal variable to be quantized; A first determination module is used to determine the integer initialization bit width corresponding to each of the integer variables to be quantized based on a probability density function and to determine the decimal initialization bit width corresponding to each of the decimal variables to be quantized based on a unified quantization method; A second determination module is used to train the agent according to the integer initialization bit width and the decimal initialization bit width, and determine the target quantization bit width output by the agent corresponding to the integer variable to be quantized and the decimal variable to be quantized respectively; The quantization module is used to simultaneously quantize the integer variable to be quantized and the decimal variable to be quantized according to the target quantization bit width to obtain a quantization result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the mixed precision quantization method for MIMO detection according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the mixed precision quantization method for MIMO detection according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the mixed precision quantization method for MIMO detection according to any one of claims 1 to 6 is implemented.

Citation Information

Patent Citations

  • Low-precision ADC quantization bit number distribution method based on deep reinforcement learning in non-cellular large-scale distributed MIMO system

    CN113660062A

  • Automatic mixing precision quantification method and device

    CN114897098A

  • Quantification method and device for MIMO detector and storage medium

    CN117749305A

  • Automatic mixed-precision quantization method and apparatus

    WO2023236609A1