A random user scheduling method based on dynamic adjustment of signal-to-interference-noise ratio threshold

By using a random user scheduling method that dynamically adjusts the signal-to-interference-plus-noise ratio (SINR) threshold, the problem of unreasonable allocation of communication resources in large-scale cellular networks is solved. This method optimizes user scheduling, shortens communication time, improves learning efficiency, and adapts to the dynamic changes in the actual network environment.

CN115397029BActive Publication Date: 2024-11-19NANTONG NORMAL COLLEGE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210974785.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-15
Publication Date
2024-11-19
Estimated Expiration
2042-08-15

AI Technical Summary

Technical Problem

In large-scale cellular networks, existing technologies have failed to effectively establish a close relationship between communication performance and learning performance, especially in the relationship between user scheduling strategies and communication time. Furthermore, they have failed to consider the differences in the importance of model updates in different rounds, resulting in unreasonable allocation of communication resources and difficulty in meeting the needs of model data upload.

Method used

A random user scheduling method based on dynamic adjustment of the signal-to-interference-plus-noise ratio (SINR) threshold is adopted. By measuring the instantaneous SINR of users, the base station dynamically adjusts the SINR threshold, selects appropriate users to participate in learning, and sends NACK signals using the transport layer protocol to optimize user scheduling. The threshold is gradually reduced to increase the number of participating users and improve learning performance.

Benefits of technology

Without significantly reducing convergence performance, it greatly shortens communication time, optimizes the use of communication resources, improves learning efficiency, and adapts to the dynamic changes in the actual network environment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115397029B_ABST
    Figure CN115397029B_ABST
Patent Text Reader

Abstract

The present invention provides a random user scheduling method based on dynamic adjustment of signal to interference noise ratio threshold, which belongs to the field of wireless communication, including: measuring the instantaneous signal to interference noise ratio of the user #imgabs0# base station by solving the problem, obtaining the optimal threshold adjustment interval M * , and the corresponding T / M * The optimized threshold value #imgabs1# is given by the base station, where T is the communication round. The base station starts federated learning, sets the initial round t=0, and sends the random initialization parameter w0 to N users. The base station receives the updated parameter #imgabs2# sent by user n and compares the parameter #imgabs3# with the instantaneous signal-to-interference-and-noise ratio #imgabs4#. When #imgabs5# is greater than #imgabs6#, the parameter #imgabs7# is included in the global parameter update, and the transport layer protocol is used to send a "NACK" signal to the unselected users. When the communication round reaches T, the base station outputs the optimal global model parameter w t This method adopts the strategy of gradually reducing the SINR threshold, which greatly shortens the communication time without significantly reducing the convergence performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of wireless communications, and in particular relates to a random user scheduling method based on dynamic adjustment of a signal to interference noise ratio threshold. Background Art

[0002] With the rapid development of IoT applications, the computing and storage capabilities of intelligent devices are constantly improving, and massive amounts of data are generated every minute in modern distributed network systems. Therefore, machine learning technology driven by big data has received widespread attention in recent years. However, due to people's attention to privacy issues and network limitations, centralized data storage and analysis methods based entirely on the cloud are no longer applicable. Especially for emergency applications that are sensitive to latency, the delay in data collection and distribution caused by long-distance communication has become an urgent problem to be solved. In this context, McMahand et al. proposed a new federated learning paradigm, advocating local and distributed learning on edge nodes. The core idea is to organize distributed devices to train data stored locally, and then send the trained model to the coordination server for aggregation, and then obtain an updated model after aggregation. Finally, the updated model is shared with all participating users to help them perform better local training in future iterations.

[0003] However, unlike server-based environments, there are many new challenges in completing distributed machine learning on mobile devices. For example, the unreliability and limited connectivity of wireless networks, the different computing capabilities of various devices, and the possibility that local data may not be independent and identically distributed (non-IID). For the above reasons, it is not feasible to allow all devices to participate in machine learning and update the model at the same time. In order to avoid long waits for outliers caused by network, computing power, etc., a more practical approach is to select only a small number of devices to participate in each round of model training. For wireless devices, the signal-to-interference-to-noise ratio (SINR) is an important indicator that affects communication reliability. It directly determines the success probability of each model upload to the server. For large-scale cellular networks, not only the interference from other users when the model is uploaded by the user to the server, but also the interference from other base stations when the aggregated model is distributed to the user, the spatial convergence analysis becomes an interesting problem.

[0004] In the prior art, HHYang, Z.Liu, TQSQuek and HVPoor published "Scheduling Policies for Federated Learning in Wireless Networks", which considered the impact of user scheduling strategy and inter-cellular interference, and studied the performance of federated learning under three strategies: random scheduling, polling scheduling and proportional fair scheduling. The authors used random geometry tools to analyze the convergence rate, and the results showed that the SINR threshold of information decoding has a great impact on the performance of the three strategies.

[0005] In the paper "Federated learning in unreliable and resource-constrained cellular wireless networks" published by M. Salehi and E. Hossain, a federated learning algorithm suitable for unreliable and resource-constrained wireless systems was proposed. The base station randomly selects some users to participate in learning and weights the user upload parameters according to the user scheduling strategy and transmission success probability to improve fairness.

[0006] Z.Lin, X.Li, VKN Lau, Y.Gong and K.Huang published "Deploying Federated Learning in Large-Scale Cellular Networks: Spatial Convergence Analysis", which simultaneously studied the convergence performance of digital aggregation and analog aggregation in cellular networks. For large-scale networks, they proposed a new learning performance indicator, the spatial convergence rate, to describe the proportion of servers that reach convergence at a certain moment. The results show that analog aggregation has a smaller learning delay than digital aggregation. The above three methods have made different attempts at user scheduling in cellular networks, but the main problem is that there is no close correlation between communication performance and learning performance, especially the relationship between communication time and user scheduling strategy. Secondly, the importance difference of model updates in different rounds is not taken into account. Reducing user participation in the first few rounds will not significantly affect the convergence speed, but it will significantly reduce communication resource occupancy and delay.

[0007] At the same time, J.Xu and H.Wang published "Client Selection and Bandwidth Allocation in Wireless Federated Learning Networks: A Long-Term Perspective" which fully recognized this problem, but did not directly incorporate this factor into the optimization of learning performance. M.Chen, Z.Yang, W.Saad, C.Yin, HVPoor and S.Cui published "A Joint Learning and Communications Framework for Federated Learning over Wireless Networks", which proposed energy saving as a constraint, but the network model was too simple and not suitable for large-scale network scenarios.

[0008] Federated edge learning is widely used in latency-sensitive big data machine learning. Large-scale cellular network users can undertake distributed learning tasks under the federated learning paradigm, but how to efficiently schedule users in cells with limited communication resources has become a key issue that needs to be solved. Especially in the 5G environment, tens of thousands of devices can simultaneously connect to base stations to share model parameters, which poses a great challenge to the allocation of communication resources. However, in the model aggregation of federated learning, the initial stage of training does not require too much user participation. As time goes on, more and more user training data is needed to correct global parameters. Limited wireless network resources cannot meet the needs of uploading a large amount of model data. Summary of the invention

[0009] In order to overcome the above-mentioned deficiencies in the prior art, the present invention provides a random user scheduling method based on dynamic adjustment of a signal to interference and noise ratio threshold.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A random user scheduling method based on dynamic adjustment of signal to interference and noise ratio threshold (SINR) includes the following steps:

[0012] Measure the user's instantaneous signal-to-interference-noise ratio

[0013] The communication round of the base station is T, and the SINR threshold γ in the tth round of communication is obtained t ;

[0014] The base station starts federated learning, sets the initial epoch t = 0, and sends the random initialization parameter w0 to all N users;

[0015] The base station receives the updated parameters sent by user n And the parameters Instantaneous Signal-to-Interference-Noise Ratio to compare; when Greater than When Incorporate global parameter updates and use the transport layer protocol to send a “NACK” signal to non-selected users;

[0016] The update equation for the global parameter update is:

[0017]

[0018] in, hour, The success probability of model upload for user n;

[0019] When the communication rounds reach T, the base station outputs the optimal global model parameter w t , to achieve user scheduling.

[0020] Preferably, the instantaneous signal to interference and noise ratio The calculation formula is:

[0021]

[0022] Among them, σ 2 is the AWGN noise power, H n is the channel fading coefficient, which obeys an exponential distribution with a mean of 1 / μ; ∈ is the power compensation factor, R n is the distance between the user and the associated base station; α is the path loss factor, and α>2;

[0023] I n It represents the total power of interference signals received by user n when uploading data, as shown in the following formula:

[0024]

[0025] Among them, D x is the distance from the user at position x to the target base station; R x is the distance from an interfering user to its corresponding associated base station, and its location x∈Φ i,n , Φ i,n represents the set of all interfering users using the same resource block as the nth user; H x is the channel gain from the user at position x to the base station.

[0026] Preferably, the base station obtains the optimal threshold adjustment interval M by solving problem P. * , and the corresponding T / M * Optimized threshold value

[0027]

[0028] Among them, T is the communication round, M is the threshold adjustment interval, τ is the sequence number of the threshold selected at equal intervals, and γ t is the SINR threshold in the tth round of communication.

[0029] Preferably, the base station uses a genetic algorithm to solve problem P, specifically:

[0030] P:

[0031]

[0032] γ min ≤γ τ+1 ≤γ τ ≤γ max

[0033]

[0034] Where T is the communication round, Θ is the learning performance, and N is the number of users; [γ min , γ max ] is the signal-to-interference-to-noise ratio range that the base station can receive the signal; q is the probability of successful signal transmission, S is the number of parameter quantization bits, γ t is the signal-to-interference-noise ratio threshold at time τ; is the signal transmission success probability of user n at time τ, M is the threshold adjustment interval; Z is a positive integer.

[0035] Preferably, the The calculation formula is:

[0036] (1) For marginal users:

[0037]

[0038] Among them, r N is the distance from the Nth user to the base station, is the marginal pdf distribution of the Nth user;

[0039] As shown below:

[0040]

[0041] and

[0042]

[0043] Where λ is the user density, Represents the Nth user at distance rN Laplace transform of the interference power when ;

[0044] Calculated by the following formula:

[0045]

[0046] Among them, e, q, and v are all temporary variables;

[0047] (2) For non-marginal users:

[0048]

[0049] in, is the joint distribution of user distances, as shown below:

[0050]

[0051] The first term of the integral is:

[0052]

[0053]

[0054] Preferably, after the base station sends the random initialization parameter w0 to all N users, the N users perform E local iterations in parallel to update the learning model parameters, and finally obtain the local parameters of user n. User n will update the parameters Send to the base station.

[0055] Preferably, the update equation for updating the learning model parameters is

[0056] in, are local model parameters, is the gradient of the local loss function, E is the communication round, and is the learning rate.

[0057] Preferably, the learning model adopts a three-layer fully connected neural network, including 200 neurons, the activation functions of the hidden layer and the output layer are hyperbolic tangent function and softmax function respectively, and the loss function is cross entropy loss.

[0058] The random user scheduling method based on dynamic adjustment of the signal to interference noise ratio threshold provided by the present invention has the following beneficial effects:

[0059] (1) In this method, the base station selects a higher SINR threshold in the initial communication round to limit low-speed users from participating in learning and shorten the communication time. Then the base station gradually lowers the SINR threshold to allow more users to participate in learning, thereby improving convergence performance.

[0060] (2) The present invention considers that the number and group of users participating in each communication round vary randomly, which is more in line with actual network conditions.

[0061] (3) The optimal user scheduling problem for optimizing communication time is summarized, and a strategy of gradually lowering the SINR threshold is adopted to significantly shorten the communication time without significantly reducing the convergence performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 This is a flow chart of a random user scheduling method based on dynamic adjustment of a signal to interference and noise ratio threshold according to Embodiment 1 of the present invention;

[0064] Figure 2 This is the experimental result of IID data distribution;

[0065] Figure 3 This is the experimental result of Non-IID data distribution. DETAILED DESCRIPTION

[0066] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.

[0067] Example 1

[0068] The present invention provides a random user scheduling method based on dynamic adjustment of signal-to-interference-plus-noise ratio (SINR) threshold for large-scale cellular network federated learning, called STUART, wherein the signal to interference plus noise ratio SINR (Signal to Interference plus Noise Ratio) refers to the ratio of the strength of the received useful signal to the strength of the received interference signal (noise and interference); it can be simply understood as "signal-to-noise ratio".

[0069] The specific implementation process of this method is as follows Figure 1 As shown, the following steps are included:

[0070] Step 1: Measure the user's instantaneous signal-to-interference-noise ratio The calculation formula is:

[0071]

[0072] Among them, σ 2 is the AWGN noise power, H n is the channel fading coefficient, which obeys an exponential distribution with a mean of 1 / μ; ∈ is the power compensation factor, R n is the distance between the user and the associated base station; α is the path loss factor, and α>2;

[0073] I n It represents the total power of interference signals received by user n when uploading data, as shown in the following formula:

[0074]

[0075] Among them, D x is the distance from the user at position x to the target base station; R x is the distance from an interfering user to its corresponding associated base station, and its location x∈Φ i,n , Φ i,n represents the set of all interfering users using the same resource block as the nth user; H x is the channel gain from the user at position x to the base station.

[0076] Step 2: The base station solves problem P and uses a genetic algorithm to obtain the optimal threshold adjustment interval M. * , and the corresponding T / M * Optimized threshold value And according to the threshold value Solve to obtain the SINR threshold γ in the tth round of communication t .

[0077]

[0078] Where T is the communication round, M is the threshold adjustment interval, τ is the sequence number of the threshold selected at equal intervals, and γ t is the SINR threshold in the tth round of communication.

[0079] Specifically, in this embodiment, the base station uses a genetic algorithm to solve problem P, specifically:

[0080] P:

[0081]

[0082] γ min ≤γ τ+1 ≤γ τ ≤γ max

[0083]

[0084] Where T is the communication round, Θ is the learning performance, and N is the number of users; [γ min , γ max ] is the signal-to-interference-to-noise ratio range that the base station can receive the signal; q is the probability of successful signal transmission, S is the number of parameter quantization bits, γ t is the signal-to-interference-noise ratio threshold at time τ; is the signal transmission success probability of user n at time τ, M is the threshold adjustment interval; Z is a positive integer.

[0085] The calculation formula is:

[0086] (1) For marginal users:

[0087]

[0088] Among them, r N is the distance from the Nth user to the base station, is the marginal pdf distribution of the Nth user;

[0089] As shown below:

[0090]

[0091] and

[0092]

[0093] Where λ is the user density, Represents the nth user at distance r N Laplace transform of the interference power when ;

[0094] Calculated by the following formula:

[0095]

[0096] Among them, e, q, and v are all temporary variables.

[0097] (2) For non-marginal users:

[0098]

[0099] in, is the joint distribution of user distances, as shown below:

[0100]

[0101] The first term of the integral is:

[0102]

[0103]

[0104] Step 3: The base station starts federated learning, sets the initial epoch t=0, and sends the random initialization parameter w0 to all N users.

[0105] Step 4: N users perform E local iterations in parallel to update the learning model parameters. The update equation is: Finally, we get the local parameters of user n

[0106] in, are local model parameters, is the gradient of the local loss function, E is the communication round, and η is the learning rate.

[0107] In this embodiment, the learning model adopts a three-layer fully connected neural network, including 200 neurons. The activation functions of the hidden layer and the output layer are the hyperbolic tangent function and the softmax function, respectively, and the loss function is the cross entropy loss. During the local training process, the present invention sets the learning rate η = 0.001, the number of communication rounds for each communication round is E = 5, and the mini-batch size is 100. All experiments were completed on a Linux server, using two NVIDIA Geforce RTX 2080Ti GPUs, and the software was MatlabR2021a.

[0108] Step 5: User n updates the parameters Send to the base station.

[0109] Step 6: The base station receives the instantaneous signal to interference noise ratio Determine whether to upload the parameters of the user Incorporate global parameter updates, specifically:

[0110] The base station receives the updated parameters sent by user n And the parameters Instantaneous Signal-to-Interference-Noise Ratio to compare; when Greater than When Incorporate global parameter updates and use the transport layer protocol to send a "NACK" signal to unselected users. NACK (Negative Acknowledgement) is a negative feedback, and the receiver notifies the sender only when it has not received data.

[0111] The update equation for global parameter update is: is the indicator function.

[0112] in, hour, γ t is the SINR threshold in the tth round of communication; is the model upload success probability of user n.

[0113] Step 7: Determine whether the communication round satisfies t<T. If so, continue to step 4. If not, end the process and the base station outputs the optimal global model parameter w t , to achieve user scheduling.

[0114] The main differences between the random user scheduling method based on dynamic adjustment of the signal to interference noise ratio threshold provided by the present invention and the traditional FedAvg are:

[0115] (1) User scheduling strategy. When the base station receives the parameters uploaded by the user, it first measures the user's instantaneous signal-to-interference-noise ratio: Then select: In addition, the method assumes that the base station can use the transport layer protocol to send a "NACK" signal to the unselected users, thereby preventing the unselected users from continuing to transmit parameters and reducing unnecessary energy consumption and interference.

[0116] (2) Model parameter aggregation. Because the probability of the uploading user being selected is not balanced, the upload success probability is: The inverse of is weighted to avoid the model parameters from shifting to users with a high probability of success.

[0117] The present invention takes the user scheduling strategy in a large-scale cellular network as the research object, takes accelerating convergence as the main goal, and obtains the optimal time-domain user scheduling algorithm under the premise of ensuring certain learning performance. The present invention assumes that the base station can accurately obtain the signal-to-interference-and-noise ratio (SINR) of each mobile user, and selects users by designing different SINR thresholds in each round to reduce communication overhead and accelerate convergence. Specifically, a higher SINR threshold is set in the initial round to allow fewer users to participate in learning. Then the SINR threshold is gradually lowered to increase the number of participating users, continuously improve learning performance, and finally reach the target accuracy. In this way, a balance is achieved between training time and training performance. This method considers that the number and group of participating users in each communication round are randomly changing, which is more in line with the actual network situation, summarizes the optimal user scheduling problem for optimizing communication time, and adopts a strategy of gradually lowering the SINR threshold to significantly shorten the communication time without significantly reducing the convergence performance.

[0118] The present invention uses the MNIST data set to test the proposed algorithm. The data set includes 60,000 28×28 pictures of handwritten numbers 0-9. These pictures are distributed to all users, with 2,400 pictures for each user. There are two distribution methods, one is independent and identically distributed (IID), in which the data owned by all users are uniformly randomly distributed. The other is non-independent and identically distributed (Non-IID), in which each user only includes data of two types of numbers. Each user uses 70% of them as a training set, 30% as a verification set, and the test set comes from 50% of all the data to ensure that the training model is fully tested.

[0119] The present invention verifies the SUART algorithm on the MNIST dataset through experiments. Figure 2 ) and No-IID( Figure 3 ) data and compared it with the benchmark scheduling strategy. The experimental results show that compared with the traditional fixed threshold strategy, the algorithm proposed in this paper can shorten the communication time by more than 1 / 3, and the convergence performance has no obvious degradation.

[0120] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.

Claims

1. A random user scheduling method based on dynamic adjustment of signal-to-interference-noise ratio threshold, characterized in that: The following steps are involved: Measure the user's instantaneous signal-to-interference-noise ratio The communication round of the base station is T, and the SINR threshold γ in the tth round of communication is obtained t ; The base station starts federated learning, sets the initial epoch t = 0, and sends the random initialization parameter w0 to all N users; The base station receives the updated parameters sent by user n And the parameters Instantaneous Signal-to-Interference-Noise Ratio to compare; when Greater than When Incorporate global parameter updates and use the transport layer protocol to send a "NACK" signal to non-selected users; The update equation for the global parameter update is: in, hour, The success probability of model upload for user n; When the communication rounds reach T, the base station outputs the optimal global model parameter w t , realize user scheduling; The instantaneous signal to interference and noise ratio The calculation formula is: Among them, σ 2 is the AWGN noise power, H n is the channel fading coefficient, which obeys an exponential distribution with a mean of 1 / μ; ∈ is the power compensation factor, R n is the distance between the user and the associated base station; α is the path loss factor, and α>2; I n It represents the total power of interference signals received by user n when uploading data, as shown in the following formula: Among them, D x is the distance from the user at position x to the target base station; R x is the distance from an interfering user to its corresponding associated base station, and its location x∈Φ i,n , Φ i,n represents the set of all interfering users using the same resource block as the nth user; H x is the channel gain from the user at position x to the base station; The base station obtains the optimal threshold adjustment interval M by solving problem P. * , and the corresponding T / M * Optimized threshold value Among them, T is the communication round, τ is the sequence number of the threshold selected at equal intervals, and γ t is the SINR threshold in the tth round of communication; The base station uses a genetic algorithm to solve problem P, specifically: Where T is the communication round, Θ is the learning performance, and N is the number of users; [γ min ,γ max ] is the signal-to-interference-to-noise ratio range that the base station can receive the signal; q is the probability of successful signal transmission, S is the number of parameter quantization bits, γ τ is the signal-to-interference-noise ratio threshold at time τ; is the probability of successful signal transmission of user n at time τ, M is the threshold adjustment interval; Z is a positive integer; (1) For marginal users: Among them, r N is the distance from the Nth user to the base station, is the marginal pdf distribution of the Nth user; As shown below: and Where λ is the user density, Represents the Nth user at distance r N Laplace transform of the interference power when ; Calculated by the following formula: Among them, e, q, and v are all temporary variables; (2) For non-marginal users: in, is the joint distribution of user distances, as shown below: The first term of the integral is: After the base station sends the random initialization parameter w0 to all N users, N users perform E local iterations in parallel to update the learning model parameters, and finally obtain the local parameters of user n User n will update the parameters Send to base station; The update equation for updating the learning model parameters is in, are local model parameters, is the gradient of the local loss function, E is the communication round, and η is the learning rate.

2. The random user scheduling method based on dynamic adjustment of signal to interference noise ratio threshold according to claim 1 is characterized in that: The learning model adopts a three-layer fully connected neural network, including 200 neurons, the activation functions of the hidden layer and the output layer are the hyperbolic tangent function and the softmax function respectively, and the loss function is the cross entropy loss.