WO2023284731A1 - Parameter setting method and apparatus, and electronic device and storage medium - Google Patents

Abstract

Provided in the embodiments of the present application are a parameter setting method and apparatus, and an electronic device and a storage medium. The method comprises: determining current performance measurement information of a communication processing model; determining a performance change trend according to the performance measurement information and historical performance measurement information; and adjusting a learning rate of the communication processing model on the basis of the performance change trend, and retraining the communication processing model according to the learning rate.

Description

Parameter setting method, device, electronic device and storage medium technical field

The present application relates to the technical field of wireless communication, for example, to a parameter setting method, device, electronic equipment and storage medium.

Background technique

With the development of wireless communication technology, network delay is gradually reduced, and communication technology is gradually becoming more intelligent and efficient. Traditional methods can no longer meet the development trend, and with the increase of hardware computing power and the maturity of artificial intelligence (AI) technology, more and more manufacturers choose to use AI technology to solve problems in wireless communication technology, for example, channel Estimation and power amplifier behavior model analysis, etc. However, in the field of AI, the training of neural networks often depends on the setting of the learning rate. For example, if the learning rate is set too high, the performance indicators of the neural network will fluctuate back and forth, resulting in failure to converge, and the neural network generated by training will lead to a decline in communication quality. If the learning rate is set too small, the performance indicators of the neural network will converge slowly, and a longer training process will lead to increased communication delays. How to set the learning rate becomes an important factor to improve communication efficiency and communication quality.

At present, there are two traditional methods to solve the learning rate setting problem:

1) Set a fixed learning rate. This setting method is often used to solve problems in fixed scenarios. However, due to the complexity of the application scenarios in wireless communication, the fixed learning rate often cannot meet the needs of multiple scenarios, resulting in neural networks generated by training. The network cannot adapt to its complex scenarios, resulting in poor communication quality.

2) Set the learning rate attenuation, initialize a larger learning rate, and attenuate the learning rate during the training process. Among them, common attenuation methods include segmental attenuation and exponential attenuation. However, since the learning rate attenuation will introduce new parameters, which increases the complexity of the communication system, resulting in uncontrollable communication quality. In addition, the learning rate attenuation process is a single attenuation. When the actual demand of the communication scene is greater than the initially set learning rate, it will cost a lot Realistically trained neural networks cannot improve communication system performance.

In view of the above problems, an adaptive learning rate setting method is urgently needed to reduce the training time of the neural network in the communication network, reduce the communication delay, and improve the communication quality.

Contents of the invention

Embodiments of the present application provide a parameter setting method, device, electronic equipment, and storage medium, so as to realize fast training of a communication processing model, reduce communication delay, and improve system communication quality.

The embodiment of the present application provides a parameter setting method, the method includes the following steps:

determining current performance metric information for the communication processing model;

determining a performance change trend according to the performance metric information and historical performance metric information;

Adjusting the learning rate of the communication processing model based on the performance change trend, and retraining the communication processing model according to the learning rate.

The embodiment of the present application also provides a parameter setting device, which includes:

The current parameter module is configured to determine the current performance measurement information of the communication processing model;

A training trend module, configured to determine a performance change trend according to the performance measurement information and historical performance measurement information;

A parameter adjustment module, configured to adjust the learning rate of the communication processing model based on the performance change trend, and retrain the communication processing model according to the learning rate.

The embodiment of the present application also provides an electronic device, the electronic device includes:

one or more processors;

memory configured to store one or more programs;

When the one or more programs are executed by the one or more processors, the one or more processors are made to implement the parameter setting method as described in any one of the embodiments of the present application.

The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the parameter setting method as described in any one of the embodiments of the present application is implemented.

In the embodiment of the present application, the performance measurement information of the communication processing model is obtained, the performance measurement information is compared with the historical performance measurement information to determine the performance change trend, the learning rate of the communication processing model is adjusted according to the performance change trend, and the learning rate is adjusted according to the learning rate Retrain the communication processing model to realize the dynamic training of the communication processing model, improve the accuracy of the learning rate of the communication processing model, and improve the communication quality of the communication system, and adjust the learning rate by using the performance change trend, which can reduce the training time of the model, thereby reducing the communication cost. The waiting time delay of the system can enhance the communication efficiency.

Description of drawings

FIG. 1 is a flow chart of a parameter setting method provided in an embodiment of the present application;

Fig. 2 is a flow chart of another parameter setting method provided by the embodiment of the present application;

Fig. 3 is a flowchart of another parameter setting method provided by the embodiment of the present application;

FIG. 4 is an example diagram of a parameter setting method provided by an embodiment of the present application;

Fig. 5 is a verification example diagram of a parameter setting method provided by the embodiment of the present application;

FIG. 6 is an example diagram of a learning rate change trend provided by an embodiment of the present application;

Fig. 7 is a comparison diagram of a neural network convergence speed provided by the embodiment of the present application;

Fig. 8 is a training example diagram of an MP model under a fixed learning rate provided by an embodiment of the present application;

Fig. 9 is a training example diagram of an MP model under a parameter setting method provided by an embodiment of the present application;

FIG. 10 is a schematic structural diagram of a communication processing model provided by an embodiment of the present application;

Fig. 11 is an example diagram of the training effect of a communication processing model provided by the embodiment of the present application;

FIG. 12 is a comparison diagram of training effects of a communication processing model provided in an embodiment of the present application;

FIG. 13 is a schematic structural diagram of another communication processing model provided by an embodiment of the present application;

Fig. 14 is an example diagram of a training effect of an LMS model provided by an embodiment of the present application;

Fig. 15 is an example diagram of the training effect of another LMS model provided by the embodiment of the present application;

Fig. 16 is a schematic structural diagram of a parameter setting device provided by an embodiment of the present application;

FIG. 17 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.

detailed description

It should be understood that the specific embodiments described here are only used to explain the present application, and are not intended to limit the present application.

In the subsequent description, use of suffixes such as 'module', 'part' or 'unit' for denoting elements is only for facilitating the description of the present application and has no specific meaning by itself. Therefore, 'module', 'part' or 'unit' may be used in combination.

Fig. 1 is a flow chart of a parameter setting method provided by the embodiment of the present application. The embodiment of the present application can be applied to the situation of deep learning model training in the communication system. The method can be executed by a parameter setting device, which can be implemented by software And/or hardware implementation, generally integrated in a centralized or communication terminal, referring to Figure 1, the method provided by the embodiment of the present application includes the following steps:

Step 110, determine the current performance measurement information of the communication processing model.

Among them, the communication processing model can be a deep learning model used to process communication parameters in the wireless communication system. The communication processing model needs to be pre-trained and generated using massive data in the wireless communication system. The accuracy of the communication processing model directly affects the communication of the wireless communication system. Quality, the communication processing model may include a power amplifier behavior model for signal processing, a digital predistortion model for improving the linearity of the power amplifier, and an adaptive equalization model for estimating the transmitted signal, etc. The performance measurement information may reflect the accuracy and processing efficiency of the communication processing model, for example, it may include the error between the output result of the communication processing model and the verification set data, and the processing speed of the communication processing model.

In the embodiment of the present application, during the training process of the communication processing model, the performance measurement information of the current state of the communication processing model may be obtained, and the acquisition method may include: inputting verification data into the communication processing model to obtain output results, and outputting The result is compared with the verification data, and the comparison result can be used as the current performance measurement information, or the timing can be started when the verification data is input into the communication processing model, and the timing can be stopped when the communication processing model generates the output result, and the timing length can be set to Performance metric information as a communication processing model.

Step 120, determine the performance change trend according to the performance measurement information and the historical performance measurement information.

Wherein, the historical performance measurement information may be the performance measurement information determined during each verification during the training process of the communication processing model, the historical performance measurement information may be all the performance measurement information determined in the history, or all the performance measurement information determined in the history The best performance value in . The performance change trend may be the comparison result between the performance measurement information and the historical performance measurement information, and may include performance improvement, performance degradation, or performance unchanged. The performance measurement information is inferior to the historical performance measurement information, and the constant performance may mean that the performance measurement information is equal to the historical performance measurement information.

For example, the performance measurement information may be compared with historical performance measurement information, and the comparison result may be used as a performance change trend. Exemplarily, a value difference between performance metric information and historical performance metric information may be determined, and a performance change trend may be determined according to the magnitude of the difference.

Step 130, adjust the learning rate of the communication processing model based on the performance change trend, and retrain the communication processing model according to the learning rate.

Wherein, the learning rate may be the rate at which data is learned in deep learning, and the learning rate may be a hyperparameter, which is used to determine whether the objective function of the communication processing model can converge to a local minimum and the convergence speed of the objective function.

In the application embodiment, the learning rate of the communication processing model can be adjusted according to the performance change trend. For example, when the performance increases, the value of the learning rate can be increased to improve the accuracy of the communication processing model; The value of the learning rate is reduced to improve the training speed of the communication processing model. After adjusting the value of the learning rate, the training data can be used to retrain the communication processing model to obtain new weights of the communication processing model.

In the embodiment of the present application, by determining the performance measurement information of the communication processing model, the performance change trend is determined according to the comparison result between the performance measurement information and the historical performance measurement information, the learning rate of the communication processing model is adjusted according to the performance change trend, and the learning rate is re-established according to the learning rate. Training the communication processing model realizes the rapid training of the communication processing model. Dynamically adjusting the learning rate through the performance change trend can improve the accuracy of the communication processing model and reduce the training time of the communication processing model, thereby reducing the waiting delay of the communication system. Enhance communication efficiency.

For example, on the basis of the foregoing application embodiments, the performance measurement information includes at least one of mean square error and normalized mean square error.

For example, Mean Squared Error (Mean Squared Error, MSE) can refer to the mean value of the square of the difference between the output result of the communication processing model and the verification data, and the calculation formula of the mean square error can be as follows:

Among them, n is the number of verification data used to verify the performance of the communication processing model, x _i is the output result, and y _i is the verification data.

In the embodiment of the present application, the normalized mean squared error (Normalized Mean Squared Error, NMSE) can be the expression of the mean squared error after transformation, converted into a dimensionless expression to become a scalar, the normalized mean squared error The calculation formula can be as follows:

Among them, n is the number of verification data used to verify the performance of the communication processing model, x _i is the output result, and y _i is the verification data.

For example, on the basis of the foregoing application embodiments, the historical performance metric information includes historical optimal performance metric information and previous training performance metric information.

In this embodiment of the present application, the historical optimal performance metric information may be the optimal information among the previously determined performance metric information, and the optimum may refer to the highest accuracy rate of the communication processing model or the fastest processing speed of the communication processing model Wait. The performance metric information of the previous training may be the performance metric information determined the previous time or several times. For example, after the performance measurement information is determined each time, the determined performance measurement information can be compared with the historical optimal performance measurement information. If the currently determined performance measurement information is better than the historical optimal performance measurement information, then the historical maximum The optimal performance metric information is replaced with the currently determined performance metric information. And, replace the previous training performance measurement information with the currently determined performance measurement information.

Fig. 2 is a flow chart of another parameter setting method provided by the embodiment of the present application. The embodiment of the present application is a refinement based on the above-mentioned embodiment of the application. Referring to Fig. 2, the method provided by the embodiment of the present application includes the following steps:

Step 210, determine the current performance measurement information of the communication processing model.

Step 220: Determine the first comparison result and the second comparison result between the performance measurement information and the historical optimal performance measurement information and the previous training performance measurement information respectively; use the first comparison result and the second comparison result as the performance change trend.

In the embodiment of the present application, the performance metric information can be compared with the historical optimal performance metric information and the previous training performance metric information respectively, and the comparison result between the performance metric information and the historical optimal performance metric information can be recorded as the first comparison result , and the comparison result between the performance measurement information and the previous training performance measurement information is recorded as the second comparison result, and the obtained first comparison result and the second comparison result can be used as a performance change trend reflecting the training state of the communication processing model.

Step 230: Decrease the value of the learning rate when the first comparison result is that the performance metric information is worse than the historical optimal performance metric information, and the second comparison result is that the performance metric information is worse than the previous training performance metric information.

For example, if the first comparison result is that the performance metric information is worse than the historical optimal performance metric information, and the second comparison result is that the performance metric information is worse than the previous training performance metric information, then it is determined that the performance of the current communication processing model is degraded after training , then reduce the value of the learning rate, so that the performance of the communication processing model will increase in the subsequent training process.

Step 240 , if the first comparison result is that the performance metric information is better than the historical optimal performance metric information, increase the value of the learning rate.

In the embodiment of the present application, if the first comparison result is that the performance metric information is better than the historical optimal performance metric information, then the performance of the current communication processing model is the highest value in history, and the value of the learning rate can be continuously increased to improve Performance of the communication processing model.

For example, if the first comparison result is that the performance metric information is worse than the historical performance metric, but the second comparison result is that the performance metric information is better than the previous training performance metric information, it means that the current learning rate is helpful to increase the performance of the communication processing model , therefore, the learning rate may not be adjusted.

Step 250, retrain the communication processing model according to the learning rate.

In the embodiment of the present application, after the value of the learning rate is adjusted, the training data may be used to retrain the communication processing model to obtain the weight of the new communication processing model.

In the embodiment of the present application, by obtaining the performance measurement information of the current communication processing model, the performance measurement information is compared with the historical optimal performance measurement information and the previous training performance measurement information to obtain the first comparison result and the second comparison result as the performance Change trend, when the first comparison result is that the performance metric information is worse than the historical optimal performance metric information and the second comparison result is that the performance metric information is worse than the previous training performance metric information, reduce the learning rate, and in the first comparison When the result is better than the historical optimal performance measurement information, increase the learning rate, rebuild and train the communication processing model according to the adjusted learning rate, realize the rapid training of the communication processing model, and dynamically adjust the learning rate through the performance change trend, which can improve The accuracy of the communication processing model reduces the training time of the communication processing model, thereby reducing the waiting delay of the communication system and enhancing communication efficiency.

Fig. 3 is a flowchart of another parameter setting method provided by the embodiment of the present application. The embodiment of the present application is a refinement based on the above-mentioned embodiment of the application. Referring to Fig. 3, the method provided by the embodiment of the present application includes the following steps:

Step 310, initialize the historical performance measurement information and the learning rate of the communication processing model.

In the embodiment of the present application, at the beginning of training the communication processing model, the historical performance measurement information and the learning rate can be initialized, and an initial value is set for the historical performance measurement information and the learning rate, and the accuracy of the initial value setting using the method of this application The performance requirements are not high, and the setting range is large, which can reduce the influence of the learning rate setting range on the communication processing model.

Step 320, determine the current performance measurement information of the communication processing model.

Step 330, determine the performance change trend according to the performance measurement information and the historical performance measurement information.

Step 340: Determine the adjustment factor corresponding to the performance change trend, and use the adjustment factor to update the value of the learning rate.

Wherein, the adjustment factor may be a weighting coefficient for adjusting the value of the learning rate.

In the embodiment of the present application, different adjustment factors can be set in advance for different performance change trends. When the performance change trend is obtained, different adjustment factors can be selected for different performance change trends, and the adjustment factor can be used to adjust the learning rate. value to be adjusted. For example, the product of the adjustment factor and the learning rate may be used as the learning rate used for training the communication processing model.

Step 350, retrain the communication processing model according to the learning rate.

Step 360: Determine the fitting relationship between the input voltage and the output voltage of the communication device according to the trained communication processing model.

For example, the communication processing model can be a communication power amplifier behavior model that processes the input voltage and output voltage of the device. This model can be used to fit the relationship between the input voltage and the output voltage of the communication device, so as to realize accurate control of the voltage of the communication device and reduce the power consumption of the device. consumption.

In the embodiment of the present application, by initializing the historical performance measurement information and the learning rate of the communication processing model, the current performance measurement information of the communication processing model is determined during the training process, and the performance change trend is determined according to the performance measurement information and historical performance measurement information, and the obtained The adjustment factor corresponding to the performance change trend, the value of the learning rate is updated according to the adjustment factor, the communication processing model is retrained using the learning rate, and the input voltage and output voltage of the communication device are fitted after the training of the communication processing model is completed to realize communication Accurate control of equipment voltage reduces voltage fluctuations, improves communication stability, and enhances communication quality.

For example, on the basis of the foregoing application embodiments, the communication processing model includes: at least one of a communication power amplifier behavior model, a digital predistortion model, and an adaptive equalization model.

In an exemplary embodiment, FIG. 4 is an example diagram of a parameter setting method provided in the embodiment of the present application. Referring to FIG. 4, it is assumed that the neural network parameter is c, the performance measurement is M, and the measurement function is M=f (c). The learning rate is μ, where μ is a real number greater than 0.

The parameter when the nth training is completed is recorded as c, the performance measure is recorded as M _n , and the learning rate is recorded as μ _n .

Record M _best as the historical optimal value of the neural network performance measure at the end of the n-1th training.

The content to be protected in this application is listed as follows:

(1) learning rate adaptive mechanism: calculate the performance measure M _n =f(c) of the network when the nth training is completed,

①If M _n is better than M _best , then let M _best =M _n , μ _n+1 =α ₁ μ _n ;

②If M _{n is} worse than M _best , and M _n is better than M _n-1 , then μ _n+1 ＝α ₂ μ _n ;

③If M _{n is} worse than M _best , and M _{n is} worse than M _n-1 , then μ _n+1 = α ₃ μ _n ;

(2) As described in (1), if the performance measure is MSE or NMSE, the superiority is equal to less than, and the inferiority is equal to greater than;

(3) As mentioned in (1), α ₁ , α ₂ , α ₃ are learning rate adjustment factors, where α ₁ ≥ 1, α ₃ ≤ 1;

(4) As described in (1), α ₁ , α ₂ , and α ₃ can be fixed values, or values that vary with the training process;

(5) As described in (1), if the maximum learning rate μ _max is configured, then μ _n+1 = min(μ _n+1 , μ _max );

(6) As described in (1), if the minimum learning rate μ _min is configured, then μ _n+1 = max(μ _n+1 , μ _min );

The above description is based on the background of the neural network, but it does not mean that the embodiment of the present application is only applicable to the neural network.

The embodiment of the present application uses the method of the embodiment of the present application to set parameters in the case of a neural network to improve the accuracy of the sinusoidal function fitting as an example to test the effectiveness of the implementation method of the present application. Referring to Figure 5, it is assumed that the input of the sinusoidal signal is x, x is 1024 numbers starting from 0 with an interval of 0.01 between two adjacent numbers. The output of the sinusoidal signal is y, y=sin(x).

A neural network model is established, the number of inputs of the model is 1, the number of outputs is 1, it contains 1 hidden layer, and the number of neurons in the hidden layer is 4. The hidden layer activation function is a tanh function, and the output layer uses a linear activation function. The parameters of the neural network are initialized, and the initialization parameters are denoted as c ₀ . Contains two parameter sets w and b, where b ₁₁ =b ₁₂ =b ₁₃ =b ₁₄ =b ₂₁ =0, w ₂₁ =-0.4228, w ₂₂ =-0.2863, w ₂₃ =-0.2793, w ₂₄ =0.0892, w ₁₁ =-0.6490, w ₁₂ =1.1812, w ₁₃ =-0.7585, w ₁₄ =-1.1096.

NMSE is used as a measure of neural network fitting performance, ie M _n =NMSE(c _n ). NMSE stands for normalized mean square error, and the smaller value is better. The expression of NMSE is:

in

is the output of the neural network to the i-th sample point x _i .

The initial value of the learning rate is set to 0.001, the maximum value of the learning rate is set to 0.1, and the minimum value is set to 0.0001.

Set fixed learning rate adjustment factors α ₁ =1.02, α ₂ =1, α ₃ =0.99.

Using the initialization parameter c ₀ to calculate the performance of neural network fitting, set M _best =M ₀ =NMSE(c ₀ )=1.0691.

After completing the above steps, start training. The training uses the Adam optimization algorithm. The intercepted part of the training records are shown in the following table:

Table 1 Training records 1-10

n=1	Mn=1.0456e+00	Mbest=1.0691e+00	μ＝0.00102(↑)
n=2	Mn=1.0278e+00	Mbest=1.0456e+00	μ＝0.00104(↑)
n=3	Mn=1.0155e+00	Mbest=1.0278e+00	μ＝0.00106(↑)
n=4	Mn=1.0070e+00	Mbest=1.0155e+00	μ＝0.00108(↑)
n=5	Mn=1.0005e+00	Mbest=1.0070e+00	μ＝0.00110(↑)
n=6	Mn=9.9527e-01	Mbest=1.0005e+00	μ＝0.00113(↑)
n=7	Mn=9.9039e-01	Mbest=9.9527e-01	μ＝0.00115(↑)
n=8	Mn=9.8562e-01	Mbest=9.9039e-01	μ＝0.00117(↑)
n=9	Mn=9.8100e-01	Mbest=9.8562e-01	μ＝0.00120(↑)

Table 1 shows the training records from n=1 to n=10. It can be seen that the performance measurement of the neural network presents a monotonous downward trend at this stage, so the learning rate continues to increase.

Table 2 Training records 271-279

n=271	Mn=3.7540e-03	Mbest=3.6499e-03	μ＝0.04725(↓)
n=272	Mn=1.5344e-02	Mbest=3.6499e-03	μ＝0.04678(↓)
n=273	Mn=1.4601e-02	Mbest=3.6499e-03	μ = 0.04678
n=274	Mn=6.4012e-03	Mbest=3.6499e-03	μ = 0.04678
n=275	Mn=3.3701e-03	Mbest=3.6499e-03	μ＝0.04772(↑)
n=276	Mn=3.0402e-03	Mbest=3.3701e-03	μ＝0.04867(↑)
n=277	Mn=3.0309e-03	Mbest=3.0402e-03	μ＝0.04965(↑)
n=278	Mn=2.7975e-03	Mbest=3.0309e-03	μ＝0.05064(↑)
n=279	Mn=2.9533e-03	Mbest=2.7975e-03	μ＝0.05013(↓)

Figure 6 is an example diagram of a learning rate change trend provided by the embodiment of the present application; see Figure 6, which shows the change trend of the learning rate during the entire training process, which increases rapidly at the beginning, and then slowly decays until the set learning rate min. Fig. 7 is a comparison diagram of the convergence speed of a neural network provided by the embodiment of the present application. Referring to Fig. 7, the learning rate adjustment method in the embodiment of the present application is compared with the fixed learning rate through the neural network, showing the proposed embodiment of the present application. The role played by the adaptive learning rate mechanism is faster than the fixed learning rate, and the final convergence performance is equivalent to the optimal fixed learning rate, and the fluctuation is small.

Exemplarily, taking the behavioral modeling of the power amplifier in the wireless communication system as an example, the relationship between the input x and the output y of the fitted power amplifier is determined, where x and y are both complex numbers. In the embodiment of the application, the communication processing The model is a memory polynomial (Memory Polynomial, MP) model, which is often used for power amplifier behavior modeling, and the expression is as follows:

Among them, Z is the output of the model, w _{k, q} are the coefficients of the model, Q represents the memory depth, and K represents the order of nonlinearity. The process of using the MP model to model the power amplifier is the process of solving the model coefficient w _k,q .

The Least Mean Square (LMS) algorithm is a common method to solve the model coefficients w _{k, q} . This algorithm is similar to the training algorithm of the neural network. It is an iterative solution algorithm based on the gradient descent method. The complex number in the general sense The principle of the LMS algorithm is as follows:

(1) Initialize w _k,q , and set the learning rate μ;

(2) The number of iterations cnt=1;

(3)n=1;

(4) Calculate the objective function

Partial derivative with respect to w _k,q

in

(5) Update parameters,

(6) n=n+1, if n≤N then return to (4);

(7) cnt=cnt+1, return (3);

The parameter setting method in the embodiment of the present application can be combined with the common LMS algorithm, and the steps are as follows:

(1) Initialize w _k,q , set learning rate μ ₁ , fixed learning rate adjustment factors α ₁ =1.3, α ₂ =1, α ₃ =0.9;

(2) Compute the performance metrics of the initial model,

(3) The number of iterations cnt=1;

(4)n=1;

(5) Calculate the objective function

Partial derivative with respect to w _k,q

in

(6) update parameters,

(7) n=n+1, if n≤N then return to (5);

(8) Calculate the performance measure of cnt iterations,

①If M _cnt is smaller than M _best , then let M _best =M _cnt , μ _cnt+1 =α ₁ μ _cnt ;

②If M _cnt is greater than M _best , and M _cnt is less than M _cnt-1 , then μ _cnt+1 = α ₂ μ _cnt ;

③If M _cnt is greater than M _best , and M _cnt is greater than M _cnt-1 , then μ _cnt+1 = α ₃ μ _cnt ;

(9) cnt=cnt+1, return to (4).

Figure 8 is a training example diagram of an MP model under a fixed learning rate provided by the embodiment of the present application. In the power amplifier modeling, the common LMS algorithm is used to set different fixed learning rates respectively. As the number of training increases, the NMSE changing circumstances. It can be seen that the final effect of different learning rates is very different. Referring to Fig. 9, in the same power amplifier modeling, the LMS algorithm combined with the parameter setting of the embodiment of the present application is used, different initial learning rates are set respectively, and the NMSE changes with the increase of the number of training times, as shown in Fig. 9 It can be seen that even if different initial learning rates are set, the training performance will quickly converge to the same level, and the convergence speed is much lower than that of MP model training with a fixed learning rate. Moreover, the setting of the initial value of the learning rate has no effect on the convergence speed of the learning rate in this application, and can reduce the influence of inaccurate setting of the initial value of the learning rate on the training of the communication processing model.

In another exemplary embodiment, when modeling the behavior of a radio frequency power amplifier in a wireless communication system, a dynamic adjustment factor may be set to adjust the learning rate, which may include the following steps:

(1) Initialize w _k,q , set learning rate μ ₁ , fixed learning rate adjustment factors α ₁ =1.3, α ₂ =1, α ₃ =0.9;

(2) Compute the performance metrics of the initial model,

(3) The number of iterations cnt=1;

(4)n=1;

(5) Calculate the objective function

Partial derivative with respect to w _k,q

in

(6) update parameters,

(7) n=n+1, if n≤N then return to (5);

(8) Calculate the performance measure of cnt iterations,

①If M _cnt is smaller than M _best , then let M _best =M _cnt , μ _cnt+1 =α ₁ μ _cnt , α ₁ =1.1α ₁ , α ₃ =1.1;

②If M _cnt is greater than M _best , and M _cnt is less than M _cnt-1 , then μ _cnt+1 = α ₂ μ _cnt ;

③If M _cnt is greater than M _best and M _cnt is greater than M _cnt-1 , then μ _cnt+1 = α ₃ μ _cnt , α ₃ =0.9α ₃ , α ₁ =1.3;

(9) cnt=cnt+1, return (4);

In the embodiment of the present application, the adjustment factors α1 and _α3 may change dynamically along with the training process _of the MP model.

On the basis of the above-mentioned embodiments, FIG. 10 is a schematic structural diagram of a communication processing model provided by the embodiment of the present application. The communication processing model shown in FIG. 10 can be used for radio frequency power amplifier modeling (Real-Valued Time-Delay Neural Networks, RVTDNN) model, RVTDNN is a real neural network that divides complex signals into real and imaginary parts to avoid the situation that traditional real neural networks cannot handle complex data. RVTDNN consists of two parts, the first part is a time delay structure, used to simulate the nonlinear characteristics of the power amplifier, and the second part is a conventional multilayer perceptron (Multilayer Perceptron, MLP) network, used to simulate the nonlinear characteristics of the power amplifier. The input x in the figure is divided into real part Real, imaginary part imag and abs three items, and each item contains 4 time delay units besides itself. These data are fed into the subsequent MLP network, which contains a hidden layer consisting of 16 nodes, and the activation function is the tanh function. The parameters of the neural network are initialized, and the initialization parameters are denoted as c ₀ . NMSE is used as a measure of neural network fitting performance, ie M _n =NMSE(c _n ). NMSE stands for normalized mean square error, and the smaller value is better. The calculation formula of NMSE is

in

is the output of the neural network for _xi .

The parameters of the neural network are initialized, and the initialization parameters are denoted as c ₀ . Then the optimal performance of the initialization history is M _best =M ₀ =NMSE(c ₀ ).

The initial value of the learning rate is set to 0.001, the maximum value of the learning rate is set to 0.1, and the minimum value is set to 0.0001.

Set fixed learning rate adjustment factors α ₁ =1.03, α ₂ =1, α ₃ =0.99.

After completing the above steps, start training. The training uses the Adam optimization algorithm. Some of the training records are intercepted as shown in the following table:

Table 3 Training records 1-9

n=1	Mn=4.3162e-02	Mbest=1.2055e+00	μ1.2055e+005e+0
n=2	Mn=1.8839e-02	Mbest=4.3162e-02	μ4.3162e-022e-0
n=3	Mn=1.1451e-02	Mbest=1.8839e-02	μ1.8839e-029e-0
n=4	Mn=8.4722e-03	Mbest=1.1451e-02	μ1.1451e-021e-0
n=5	Mn=6.2968e-03	Mbest=8.4722e-03	μ8.4722e-032e-0
n=6	Mn=4.7780e-03	Mbest=6.2968e-03	μ6.2968e-038e-0
n=7	Mn=3.7248e-03	Mbest=4.7780e-03	μ4.7780e-030e-0
n=8	Mn=2.7483e-03	Mbest=3.7248e-03	μ3.7248e-038e-0
n=9	Mn=2.3303e-03	Mbest=2.7483e-03	μ2.7483e-033e-0

Table 3 shows the training records from n=1 to n=9. It can be seen that the performance measurement of the neural network presents a monotonous downward trend at this stage, so the learning rate continues to increase.

Table 4 Training records 271-279

n=57	Mn=5.2272e-04	Mbest=3.9255e-04	μ3.9255e-045e-0
n=58	Mn=4.2117e-04	Mbest=3.9255e-04	μ3.9255e-045
n=59	Mn=3.8949e-04	Mbest=3.9255e-04	μ3.9255e-045e-0
n=60	Mn=5.3353e-04	Mbest=3.8949e-04	μ3.8949e-049e-0
n=61	Mn=3.9073e-04	Mbest=3.8949e-04	μ3.8949e-049
n=62	Mn=3.8795e-04	Mbest=3.8949e-04	μ3.8949e-049e-0
n=63	Mn=3.5915e-04	Mbest=3.8795e-04	μ3.8795e-045e-0
n=64	Mn=4.2375e-04	Mbest=3.5915e-04	μ3.5915e-045e-0
n=65	Mn=4.2229e-04	Mbest=3.5915e-04	μ3.5915e-045

Table 4 shows the training records of the RVTDNN model from n=57 to n=65. It can be seen that the neural network performance metrics fluctuate up and down at this stage, and the learning rate sometimes increases, sometimes decreases, and sometimes remains unchanged. Among them, the learning rate remains unchanged when n is equal to 58, 61 and 65, the learning rate increases when n is equal to 59, 62 and 63, and the learning rate decreases at other times. Figure 11 is an example diagram of the training effect of a communication processing model provided by the embodiment of the present application. Referring to Figure 11, the change trend of the learning rate during the entire training process increases rapidly at the beginning, and then slowly decays until the set minimum value. Fig. 12 is a comparison diagram of the training effect of a communication processing model provided by the embodiment of the present application. Fig. 12 shows the RVTDNN model applying the adaptive learning rate set by the parameter setting method of the embodiment of the present application and the training using the fixed learning rate method The comparison results show that compared with the fixed learning rate, the performance measurement information of the adaptive learning rate set by the parameter setting method provided by the embodiment of the present application converges faster, and the final convergence performance is also better.

In an exemplary implementation, FIG. 13 is a schematic structural diagram of another communication processing model provided by the embodiment of the present application. In a wireless communication system, the nonlinearity of the power amplifier of the transmitter will cause distortion of the transmitted signal, resulting in a decrease in communication performance. , in order to avoid this situation, it is necessary to use a communication processing model to linearly amplify the original signal. The communication processing model can be a digital pre-distortion (Digital Pre-Distortion, DPD) model, which allows the transmitted signal to be processed by the DPD module first. , and then fed into the power amplifier, so that the output signal of the power amplifier is a linear amplification of the original signal. The memory polynomial model is a commonly used DPD module model, and its expression is as follows:

Among them, Z is the output of the model, w _{k, q} are the coefficients of the model, Q represents the memory depth, and K represents the order of nonlinearity. The core of the DPD problem is to solve the coefficient w _k,q of the DPD model. Figure 13 shows an LMS algorithm for solving DPD coefficients with an indirect architecture. The training process of the LMS algorithm is similar to the algorithm of the neural network. The parameter setting method in the embodiment of the present application can be used. The processing process of the common LMS algorithm can be Including the following steps:

(1) Initialize w _k,q , and set the learning rate μ;

(2) The number of iterations cnt=1;

(3)n=1;

(4) Calculation error, e(n)=y(n)/G-x(n), where G is the amplification factor of the power amplifier;

(5) Update parameters, w _{k, q} = w _{k, q} -μ(x[nq]|x[nq]| ^k-1 ) ^* e;

(6) n=n+1, if n≤N then return to (4);

(7) cnt=cnt+1, return to (3).

And the processing process of the LMS algorithm combined with the parameter setting method in the embodiment of the present application may include the following steps:

(1) Initialize w _k,q , set initial learning rate μ ₁ , fixed learning rate adjustment factor

Acceleration factor ν ₁ = 1.01, ν ₃ = 0.999;

(2) Compute the performance metrics of the initial model,

(3) The number of iterations cnt=1;

(4)n=1;

(5) Calculation error, e(n)=y(n)/G-x(n);

(6) Update parameters, w _{k, q} = w _{k, q} -μ(x[nq]|x[nq]| ^k-1 ) ^* e;

(7) n=n+1, if n≤N then return to (5);

(8) Calculate the performance measure of cnt iterations,

a) If M _cnt is smaller than M _best , then set M _best = M _cnt , μ _cnt+1 = α ₁ μ _cnt , α ₁ =ν ₁ α ₁ ,

b) If M _cnt is greater than M _best and M _best is less than M _cnt-1 , then μ _cnt+1 = μ _cnt ;

c) If M _cnt is greater than M _best , and M _best is greater than M _cnt-1 , then μ _cnt+1 = α ₃ μ _cnt , α ₃ =ν ₃ α ₃ ,

(9) cnt=cnt+1, return to (4).

Figure 14 is an example diagram of the training effect of an LMS model provided by the embodiment of the present application. Referring to Figure 14, it shows that in the digital predistortion model, the common LMS algorithm is used to set different fixed learning rates respectively. In the case of increasing NMSE changes, it can be seen that the final effects of different learning rates are very different. The learning rate trained by related technologies is extremely unstable, which can easily lead to distortion of the signal sent by the communication system and reduce the communication quality. Fig. 15 is an example diagram of the training effect of another LMS model provided by the embodiment of the present application. Referring to Fig. 15, the LMS algorithm combined with the parameter setting method in the embodiment of the present application is used in the same digital predistortion model. Different initial learning rates are set respectively, and the NMSE changes with the increase of training times. It can be seen that even if different initial learning rates are set, the training performance will quickly converge to the same level. In the embodiment of the present application, the setting of the initial learning rate is insensitive, which is convenient for the user to train the digital predistortion model. _In this embodiment, the adjustment factors α1 and _α3 change dynamically along with the training process.

In another exemplary embodiment, in a wireless communication system, a signal sent by a transmitter will be affected by multipath propagation when passing through a wireless channel, and an equalization module needs to be used on the receiver side to restore the original signal. Assuming that the transmitted signal is x, the received signal is y, and the channel impulse response is h, the expression of the process is:

Among them, equalization is the process of recovering x from y, and the equalization can adopt a linear model, for example:

Among them, the coefficient w of the model can be solved using the LMS algorithm. The equalization can also be the RVTDNN model in the above embodiment, and the coefficients of the model can be solved using a back propagation (Back Propagation, BP) algorithm. No matter what kind of model is used, the setting of the learning rate can adopt the parameter setting method proposed in the embodiment of the present application.

Fig. 16 is a schematic structural diagram of a parameter setting device provided by an embodiment of the present application, which can execute the parameter setting method provided by any embodiment of the present application, and has corresponding functional modules and beneficial effects for executing the method. The device can be controlled by software and/or Hardware implementation, including: a current parameter module 401 , a training trend module 402 and a parameter adjustment module 403 .

The current parameter module 401 is configured to determine the current performance measurement information of the communication processing model.

The training trend module 402 is configured to determine a performance change trend according to the performance metric information and historical performance metric information.

The parameter adjustment module 403 is configured to adjust the learning rate of the communication processing model based on the performance change trend, and retrain the communication processing model according to the learning rate.

In the embodiment of the present application, the performance measurement information of the communication processing model is determined by the current parameter module, the training trend module determines the performance change trend according to the comparison result between the performance measurement information and the historical performance measurement information, and the parameter adjustment module adjusts the communication processing model according to the performance change trend Learning rate, and retraining the communication processing model according to the learning rate, realizes the rapid training of the communication processing model, and dynamically adjusts the learning rate through the performance change trend, which can improve the accuracy of the communication processing model and reduce the training time of the communication processing model, thus Reducing the waiting delay of the communication system can enhance the communication efficiency.

For example, on the basis of the above application embodiments, the performance measurement information in the current parameter module 301 includes at least one of mean square error and normalized mean square error.

For example, on the basis of the foregoing application embodiments, the historical performance metric information in the training trend module 302 includes historical optimal performance metric information and previous training performance metric information.

For example, on the basis of the above-mentioned application embodiments, the training trend module 302 includes:

The comparison execution unit is configured to determine a first comparison result and a second comparison result between the performance metric information and the historical optimal performance metric information and the previous training performance metric information respectively.

A trend determination unit is configured to use the first comparison result and the second comparison result as the performance change trend.

For example, on the basis of the above application embodiments, the parameter adjustment module 303 includes:

The first processing unit is configured to, when the first comparison result is that the performance metric information is worse than the historical optimal performance metric information, and the second comparison result is that the performance metric information is worse than the previous time In the case of training performance metric information, reduce the value of the learning rate.

The second processing unit is configured to increase the value of the learning rate when the first comparison result is that the performance metric information is better than the historical optimal performance metric information.

For example, on the basis of the above-mentioned application embodiments, the parameter adjustment module 303 in the device further includes:

The factor adjustment unit is configured to determine an adjustment factor corresponding to the performance change trend, and use the adjustment factor to update the value of the learning rate.

For example, on the basis of the above application embodiments, the device further includes:

A parameter initialization module, configured to initialize the historical performance measurement information and the learning rate of the communication processing model.

For example, on the basis of the above application embodiments, the device further includes:

The model use module is configured to determine the fitting relationship between the input voltage and the output voltage of the communication device according to the trained communication processing model.

For example, on the basis of the foregoing application embodiments, the communication processing model in the device includes: at least one of a communication power amplifier behavior model, a predistortion model, and an adaptive equalization model.

Figure 17 is a schematic structural diagram of an electronic device provided by an embodiment of the present application, the electronic device includes a processor 50, a memory 51, an input device 52 and an output device 53; the number of processors 50 in the electronic device can be one or more In FIG. 17, a processor 50 is taken as an example; the processor 50, memory 51, input device 52 and output device 53 in the electronic device can be connected through a bus or in other ways. In FIG. 17, the connection through a bus is taken as an example.

Memory 51, as a computer-readable storage medium, can be set to store software programs, computer-executable programs and modules, such as the modules corresponding to the parameter setting device in the embodiment of the present application (current parameter module 401, training trend module 402 and parameter adjustment module 403). The processor 50 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 51 , that is, realizes the above parameter setting method. The computer readable storage medium may be a non-transitory computer readable storage medium.

The memory 51 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required by a function; the data storage area may store data created according to the use of the electronic device, and the like. In addition, the memory 51 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage devices. In some examples, memory 51 may include memory located remotely relative to processor 50, and these remote memories may be connected to the electronic device through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The input device 52 can be configured to receive input numbers or character information, and generate key signal input related to user settings and function control of the electronic device. The output device 53 may include a display device such as a display screen.

The embodiment of the present application also provides a storage medium containing computer-executable instructions, the computer-executable instructions are configured to execute a parameter setting method when executed by a computer processor, the method comprising:

determining current performance metric information for the communication processing model;

determining a performance change trend according to the performance metric information and historical performance metric information;

Adjusting the learning rate of the communication processing model based on the performance change trend, and retraining the communication processing model according to the learning rate.

Through the above descriptions about the implementation manners, those skilled in the art can clearly understand that the present application can be realized by software and necessary general-purpose hardware, and of course it can also be realized by hardware. Based on this understanding, the essence of the technical solution of this application or the part that contributes to related technologies can be embodied in the form of software products, and the computer software products can be stored in computer-readable storage media, such as computer floppy disks, Read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disc, etc., including several instructions to make a computer device (which can be a personal computer, A server, or a network device, etc.) executes the methods described in multiple embodiments of the present application.

It is worth noting that, in the embodiment of the above-mentioned device, the multiple units and modules included are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be realized; in addition, multiple The specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the protection scope of the present application.

Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, the functional modules/units in the system, and the device can be implemented as software, firmware, hardware, and an appropriate combination thereof.

In a hardware implementation, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be composed of several physical components. Components cooperate to execute. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit . Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). As known to those of ordinary skill in the art, the term computer storage media includes both volatile and nonvolatile media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, program modules, or other data. permanent, removable and non-removable media. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cartridges, tape, magnetic disk storage or other magnetic storage devices, or can Any other medium used to store desired information and which can be accessed by a computer. In addition, as is well known to those of ordinary skill in the art, communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

The embodiment of the present application realizes the dynamic update of the learning rate, improves the accuracy of the communication processing model, can reduce the training time of the communication processing model, thereby reducing the waiting time of the communication system, enhancing communication efficiency, and improving communication quality.

Claims (13)

1. A parameter setting method, comprising:

   determining current performance metric information for the communication processing model;

   determining a performance change trend according to the performance metric information and historical performance metric information;

   Adjusting the learning rate of the communication processing model based on the performance change trend, and retraining the communication processing model according to the learning rate.
2. The method according to claim 1, wherein the performance measurement information includes at least one of mean square error and normalized mean square error.
3. The method according to claim 1, wherein the historical performance metric information includes historical optimal performance metric information and previous training performance metric information.
4. The method according to claim 3, wherein said determining the performance change trend according to the performance metric information and historical performance metric information comprises:

   determining a first comparison result and a second comparison result between the performance metric information and the historical optimal performance metric information and the previous training performance metric information;

   The first comparison result and the second comparison result are used as the performance change trend.
5. The method according to claim 4, wherein the adjusting the learning rate of the communication processing model based on the performance change trend comprises at least one of the following:

   In response to determining that the first comparison result is that the performance metric information is inferior to the historical optimal performance metric information, and that the second comparison result is that the performance metric information is inferior to the previous training performance metric information, Decrease the value of the learning rate;

   In response to determining that the first comparison result is that the performance metric information is better than the historical optimal performance metric information, increasing the value of the learning rate.
6. The method according to claim 1, wherein the adjusting the learning rate of the communication processing model based on the performance variation trend comprises:

   An adjustment factor corresponding to the performance change trend is determined, and the value of the learning rate is updated using the adjustment factor.
7. The method according to claim 6, wherein the adjustment factors corresponding to different performance change trends are different.
8. The method according to claim 1, further comprising:

   Initializing the historical performance metric information and the learning rate of the communication processing model.
9. The method according to claim 1, further comprising:

   A fitting relationship between the input voltage and the output voltage of the communication device is determined according to the communication processing model that has been trained.
10. The method according to claim 1, wherein the communication processing model comprises: at least one of a communication power amplifier behavior model, a digital predistortion model, and an adaptive equalization model.
11. A parameter setting device, comprising:

    The current parameter module is configured to determine the current performance measurement information of the communication processing model;

    A training trend module, configured to determine a performance change trend according to the performance measurement information and historical performance measurement information;

    A parameter adjustment module, configured to adjust the learning rate of the communication processing model based on the performance change trend, and retrain the communication processing model according to the learning rate.
12. An electronic device comprising:

    one or more processors;

    memory configured to store one or more programs;

    When the one or more programs are executed by the one or more processors, the one or more processors are made to implement the parameter setting method according to any one of claims 1-10.
13. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the parameter setting method according to any one of claims 1-10 is implemented.

Images