CN116418687A - Parameter optimization and resource allocation method for low-energy wireless federated learning system - Google Patents

Abstract

The invention discloses a parameter optimization and resource allocation method of a low-energy-consumption wireless federal learning system, which comprises the steps of constructing a low-energy-consumption wireless federal learning system model and establishing a federal learning flow; and establishing a joint parameter optimization and resource allocation problem by taking the minimum system model energy consumption as an optimization target, splitting the joint parameter optimization and resource allocation problem into an aggregation interval optimization sub-problem, calculating to obtain an optimal aggregation interval and frequency optimization sub-problem, calculating to obtain an optimal calculation frequency and bandwidth allocation optimization sub-problem, calculating to obtain an optimal bandwidth allocation and transmission power optimization sub-problem, and calculating to obtain the optimal transmission power of the intelligent terminal and the edge server. The method carries out joint optimization on a plurality of variables of wireless communication and model training, minimizes the system energy consumption in the federal learning process under the limit of model training performance, and can obviously reduce the system energy consumption.

Description

Parameter optimization and resource allocation method for low-energy wireless federated learning system

technical field

The invention belongs to the technical field of wireless communication, and in particular relates to a parameter optimization and resource allocation method of a low-energy wireless federated learning system.

Background technique

Compared with the traditional centralized machine learning model training method, federated learning can make full use of the local data and computing resources of distributed intelligent terminals, and effectively complete the distribution of machine learning models under the premise of protecting the data privacy of intelligent terminals and saving communication resources. style training. However, in hybrid wireless heterogeneous networks, the working time of smart terminals is often limited by their limited energy reserves. Therefore, how to effectively reduce system energy consumption while ensuring model training performance has become one of the urgent problems to be solved in federated learning.

Contents of the invention

Aiming at the above problems, the present invention provides a method for parameter optimization and resource allocation of a low-energy wireless federated learning system. In the wireless federated learning system, energy consumption mainly includes calculation energy consumption for model training and communication energy consumption for model parameter transmission. , a parameter optimization and resource allocation method to reduce energy consumption is designed.

Technical scheme of the present invention is as follows:

A method for parameter optimization and resource allocation of a low-energy wireless federated learning system, comprising the following steps:

Build a low-energy wireless federated learning system model, including a cloud server, N edge servers and M smart terminals, and establish a federated learning process based on the low-energy wireless federated learning system model;

With the optimization goal of minimizing the energy consumption of the system model, limit the number of smart terminal local model training related to the aggregation interval, the calculation frequency of the smart terminal, the system bandwidth allocation, and the transmission power of the smart terminal and the edge server, and establish joint parameter optimization and resource allocation. question,

Split the joint parameter optimization and resource allocation problem into aggregation interval optimization sub-problem and calculate the optimal aggregation interval, frequency optimization sub-problem and calculate the optimal computing frequency, bandwidth allocation optimization sub-problem and calculate the optimal bandwidth allocation and transmission The power optimization sub-problem is calculated to obtain the optimal transmission power of the smart terminal and the edge server.

A further technical solution of the present invention is: establishing a federated learning process based on a low-energy wireless federated learning system model, specifically including the following steps:

S1. The cloud server broadcasts the global model and its current parameter values to all smart terminals through the edge server;

S2. The smart terminal uses local data to iteratively train the received global model. After κ ₁ local iterative training is completed, the smart terminal uploads the updated global model parameters after training, that is, the local model parameters, to the edge server;

S3. After all the associated smart terminals upload the parameters, the edge server aggregates the received model parameters into the parameters of the edge model, and distributes them to all the smart terminals associated with the edge server through multicast;

S4. After performing κ ₂ iterations on S2 and S3, the edge server uploads the parameters of the edge model to the cloud server;

S5. After all the edge servers upload the edge model parameters, the cloud server aggregates the received model parameters, and updates the global model according to the aggregated model parameters;

S6. Iteratively execute S1 to S5 until the global model converges or reaches a preset precision value.

The further technical solution of the present invention is: the method also includes calculating the computing delay and computing energy consumption of the smart terminal in the local iterative training based on the federated learning process, the rate at which the smart terminal uploads the global model parameters to the edge server, and the group of edge servers broadcast rate, including:

其中，δ是智能终端的能量消耗系数；The calculation delay of smart terminal m in a local iterative training is: T _m,cop =v _m CD _m /f _m , where v _m is the number of CPU rounds required by the terminal to process unit bit sample data, and C is the sample data The bit size of D _m is the number of samples of the terminal, f _m is the calculation frequency of the intelligent terminal, and the corresponding calculation energy consumption of the intelligent terminal is expressed as:

Among them, δ is the energy consumption coefficient of the smart terminal;

其中，B_n为边缘服务器n的带宽，p_m为智能终端m的传输功率，g_n,m为信道增益，N₀为加性噪声功率；智能终端的参数上传时间为T_m,com＝W/R_m，其中W为上传模型参数的比特大小，智能终端的通信能耗为E_m,com＝p_mT_m,com；The rate at which smart terminal m uploads local model parameters to edge server n is:

Among them, B _n is the bandwidth of the edge server n, p _m is the transmission power of the intelligent terminal m, g _n,m is the channel gain, N ₀ is the additive noise power; the parameter upload time of the intelligent terminal is T _m,com =W /R _m , where W is the bit size of the uploaded model parameters, and the communication energy consumption of the smart terminal is E _{m, com} = p _m T _{m, com} ;

其中，p_n为边缘服务器n的传输功率，C_n为与边缘服务器n相关联的所有智能终端的索引集合，边缘服务器的下行链路组播时延为T_n,com＝W/R_n，边缘服务器的通信能耗为E_n,com＝p_nT_n,com。The multicast rate of edge server n is:

Among them, p _n is the transmission power of the edge server n, C _n is the index set of all intelligent terminals associated with the edge server n, and the downlink multicast delay of the edge server is T _{n, com} = W/R _n , The communication energy consumption of the edge server is E _n,com =p _n T _n,com .

The further technical scheme of the present invention is: the concrete expression of system model energy consumption is:

A further technical solution of the present invention is: the specific expression of the joint parameter optimization and resource allocation problem is:

表示聚合间隔集合，/>

表示计算频率集合，

表示带宽集合，/>

表示传输功率集合，κ_glob为一轮全局模型迭代中的本地模型训练次数，f_m,max为终端最大计算频率，/>

和/>

为智能终端的最大计算时间和最大通信时间，T_n,max为边缘服务器组播的最大时延，B_total为系统总带宽，B_min和B_max分别为边缘服务器的最小带宽和最大带宽，p_m,max和p_n,max分别为智能终端的最大传输功率和边缘服务器的最大传输功率，E_m,max和E_n,max分别为智能终端的最大能耗和边缘服务器的最大能耗。in,

Represents a collection of aggregation intervals, />

Represents the set of computing frequencies,

Represents a bandwidth collection, />

Indicates the set of transmission power, κ _glob is the number of local model training in one round of global model iteration, f _m,max is the maximum calculation frequency of the terminal, />

and />

is the maximum calculation time and maximum communication time of the intelligent terminal, T _n,max is the maximum delay of the edge server multicast, B _total is the total system bandwidth, B _min and B _max are the minimum bandwidth and maximum bandwidth of the edge server respectively, p _m,max and p _n,max are the maximum transmission power of the intelligent terminal and the maximum transmission power of the edge server, respectively, and E _m,max and E _n,max are the maximum energy consumption of the intelligent terminal and the maximum energy consumption of the edge server, respectively.

The further technical solution of the present invention is: split the joint parameter optimization and resource allocation problem into aggregation interval optimization sub-problems and calculate the optimal aggregation interval, the specific expression is:

The further technical solution of the present invention is: split the joint parameter optimization and resource allocation problem into frequency optimization sub-problems and calculate the optimal calculation frequency, the specific expression is:

The further technical solution of the present invention is: split the joint parameter optimization and resource allocation problem into transmission power optimization sub-problems and calculate the optimal transmission power of the smart terminal and the edge server, the specific expression is:

A further technical solution of the present invention is: splitting the joint parameter optimization and resource allocation problem into bandwidth allocation optimization sub-problems and calculating the optimal bandwidth allocation, wherein the optimal bandwidth allocation is realized by using a convex optimization toolbox.

The parameter optimization and resource allocation method of a low-energy wireless federated learning system provided by the present invention constructs a low-energy wireless federated learning system model, and establishes a federated learning process based on the low-energy wireless federated learning system model; to minimize the energy consumption of the system model To optimize the goal, establish a joint parameter optimization and resource allocation problem, split the joint parameter optimization and resource allocation problem into aggregation interval optimization sub-problems and calculate the optimal aggregation interval, frequency optimization sub-problems and calculate the optimal computing frequency and bandwidth Assign optimization sub-problems and calculate the optimal bandwidth allocation and transmission power optimization sub-problems and calculate the optimal transmission power of smart terminals and edge servers. The invention jointly optimizes multiple variables of wireless communication and model training, and minimizes the system energy consumption in the federated learning process under the limitation of model training performance. Compared with other schemes, the method of the present invention is significantly better than the comparison scheme. Can significantly reduce system energy consumption.

Description of drawings

FIG. 1 is a schematic flow diagram of a method for parameter optimization and resource allocation of a low-energy wireless federated learning system in an embodiment of the present invention;

Fig. 2 is a schematic structural diagram of a wireless hierarchical federated learning system model in an embodiment of the present invention;

Fig. 3 is a comparison chart of the variation trend of system energy consumption with the number of smart terminals associated with the edge server under different optimization schemes in the embodiment of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, but not to limit the present invention. In addition, it should be noted that, for the convenience of description, only some structures related to the present invention are shown in the drawings but not all structures.

Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processing, many of the steps may be performed in parallel, concurrently, or simultaneously. Additionally, the order of steps may be rearranged. The process may be terminated when its operations are complete, but may also have additional steps not included in the figure. The processing may correspond to a method, function, procedure, subroutine, subroutine, or the like.

As shown in Figure 1, a method for parameter optimization and resource allocation of a low-energy wireless federated learning system in an embodiment includes the following steps:

Build a low-energy wireless federated learning system model, including a cloud server, N edge servers and M smart terminals, and establish a federated learning process based on the low-energy wireless federated learning system model;

With the optimization goal of minimizing the energy consumption of the system model, limit the number of smart terminal local model training related to the aggregation interval, the calculation frequency of the smart terminal, the system bandwidth allocation, and the transmission power of the smart terminal and the edge server, and establish joint parameter optimization and resource allocation. question;

Split the joint parameter optimization and resource allocation problem into aggregation interval optimization sub-problem and calculate the optimal aggregation interval, frequency optimization sub-problem and calculate the optimal computing frequency, bandwidth allocation optimization sub-problem and calculate the optimal bandwidth allocation and transmission The power optimization sub-problem is calculated to obtain the optimal transmission power of the smart terminal and the edge server.

In the specific implementation process, as shown in FIG. 2 , a wireless hierarchical federated learning system provided by the embodiment includes a cloud server, N edge servers and M intelligent terminals. The cloud server and each edge server are connected through a high-performance wired backhaul link, and the edge server and its associated intelligent terminals are connected with a wireless communication link.

Preferably, the federated learning process is established based on the low-energy wireless federated learning system model, which specifically includes the following steps:

S1. The cloud server broadcasts the global model and its current parameter values to all smart terminals through the edge server;

S2. The smart terminal uses local data to iteratively train the received global model. After κ ₁ local iterative training is completed, the smart terminal uploads the updated global model parameters after training, that is, the local model parameters, to the edge server;

S3. After all the associated smart terminals upload the parameters, the edge server aggregates the received model parameters into the parameters of the edge model, and distributes them to all the smart terminals associated with the edge server through multicast;

S4. After performing κ ₂ iterations on S2 and S3, the edge server uploads the parameters of the edge model to the cloud server;

S5. After all the edge servers upload the edge model parameters, the cloud server aggregates the received model parameters, and updates the global model according to the aggregated model parameters;

S6. Iteratively execute S1 to S5 until the global model converges or reaches a preset precision value. So far, the machine learning model training process of the federated learning system is over.

During the specific implementation process, the calculation delay and energy consumption of the smart terminal in local iterative training, the rate at which the smart terminal uploads global model parameters to the edge server, and the multicast rate of the edge server are calculated based on the federated learning process, including:

其中，δ是智能终端的能量消耗系数；The calculation delay of smart terminal m in a local iterative training is: T _m,cop =v _m CD _m /f _m , where v _m is the number of CPU rounds required by the terminal to process unit bit sample data, and C is the sample data The bit size of D _m is the number of samples of the terminal, f _m is the calculation frequency of the intelligent terminal, and the corresponding calculation energy consumption of the intelligent terminal is expressed as:

Among them, δ is the energy consumption coefficient of the smart terminal;

其中，B_n为边缘服务器n的带宽，p_m为智能终端m的传输功率，g_n,m为信道增益，N₀为加性噪声功率；智能终端的参数上传时间为T_m,com＝W/R_m，其中W为上传模型参数的比特大小，智能终端的通信能耗为E_m,com＝p_mT_m,com；The rate at which smart terminal m uploads local model parameters to edge server n is:

Among them, B _n is the bandwidth of the edge server n, p _m is the transmission power of the intelligent terminal m, g _n,m is the channel gain, N ₀ is the additive noise power; the parameter upload time of the intelligent terminal is T _m,com =W /R _m , where W is the bit size of the uploaded model parameters, and the communication energy consumption of the smart terminal is E _{m, com} = p _m T _{m, com} ;

其中，p_n为边缘服务器n的传输功率，/>

为与边缘服务器n相关联的所有智能终端的索引集合，边缘服务器的下行链路组播时延为T_n,com＝W/R_n，边缘服务器的通信能耗为E_n,com＝p_nT_n,com。The multicast rate of edge server n is:

Among them, p _n is the transmission power of edge server n, />

is the index set of all intelligent terminals associated with edge server n, the downlink multicast delay of the edge server is T _n,com =W/R _n , and the communication energy consumption of the edge server is E _n,com =p _n T _{n, com} .

In addition, the wireless communication link in the system adopts the "hybrid frequency-time bandwidth allocation" scheme. The total bandwidth of the system is allocated to each edge server by frequency division, and each intelligent terminal uses time division in turn to use the bandwidth allocated by its edge server. bandwidth.

Further, the specific expression of the energy consumption of the system model is:

计算频率/>

带宽分配/>

以及传输功率/>

The embodiment aims to minimize the system energy consumption in the above definition, and jointly performs parameter optimization and resource allocation, specifically involving the aggregation interval

Calculation frequency />

bandwidth allocation/>

and transmission power/>

The joint parameter optimization and resource allocation problem can be modeled as:

表示聚合间隔集合，/>

表示计算频率集合，

表示带宽集合，/>

表示传输功率集合，κ_glob为一轮全局模型迭代中的本地模型训练次数，f_m,max为终端最大计算频率，/>

和/>

为智能终端的最大计算时间和最大通信时间，T_n,max为边缘服务器组播的最大时延，B_total为系统总带宽，B_min和B_max分别为边缘服务器的最小带宽和最大带宽，p_m,max和p_n,max分别为智能终端的最大传输功率和边缘服务器的最大传输功率，E_m,max和E_n,max分别为智能终端的最大能耗和边缘服务器的最大能耗。in,

Represents a collection of aggregation intervals, />

Represents the set of computing frequencies,

Represents a bandwidth collection, />

Indicates the set of transmission power, κ _glob is the number of local model training in one round of global model iteration, f _m,max is the maximum calculation frequency of the terminal, />

and />

is the maximum calculation time and maximum communication time of the intelligent terminal, T _n,max is the maximum delay of the edge server multicast, B _total is the total system bandwidth, B _min and B _max are the minimum bandwidth and maximum bandwidth of the edge server respectively, p _m,max and p _n,max are the maximum transmission power of the intelligent terminal and the maximum transmission power of the edge server, respectively, and E _m,max and E _n,max are the maximum energy consumption of the intelligent terminal and the maximum energy consumption of the edge server, respectively.

Preferably, the embodiment splits the above-mentioned joint optimization problem into four sub-problems: aggregation interval optimization, calculation frequency optimization, bandwidth allocation optimization, and transmission power optimization:

Among them, the joint parameter optimization and resource allocation problem is split into aggregation interval optimization sub-problems and the optimal aggregation interval is calculated. The specific expression is:

Split the joint parameter optimization and resource allocation problem into frequency optimization sub-problems and calculate the optimal computing frequency, the specific expression is:

The problem of joint parameter optimization and resource allocation is divided into transmission power optimization sub-problems and the optimal transmission power of smart terminals and edge servers is calculated. The specific expression is:

In addition, the problem of joint parameter optimization and resource allocation is split into bandwidth allocation optimization sub-problems and the optimal bandwidth allocation is calculated. The optimal bandwidth allocation is realized by convex optimization toolbox.

The process of parameter optimization and resource allocation method based on low-energy wireless federated learning is summarized in Algorithm 1 as follows:

In order to better reflect the effectiveness of the present invention, a simulation experiment is carried out in the embodiment. Figure 3 shows the variation trend of system energy consumption with the number of smart terminals associated with the edge server under different optimization schemes. In the experiment, the total number M of intelligent terminals in the system is set to 2000 and 3000 respectively. It can be observed that the parameter optimization and resource allocation method proposed by the present invention can be consistent with the exhaustive search algorithm of the joint optimization problem, and at the same time, it is significantly better than the comparison scheme, and can significantly reduce system energy consumption.

It can be seen from the embodiments that a method for parameter optimization and resource allocation of a low-energy wireless federated learning system provided by the present invention constructs a low-energy wireless federated learning system model, and establishes a federated learning process based on the low-energy wireless federated learning system model; With the optimization goal of minimizing the energy consumption of the system model, a joint parameter optimization and resource allocation problem is established, and the joint parameter optimization and resource allocation problem is split into aggregation interval optimization sub-problems, and the optimal aggregation interval and frequency optimization sub-problems are calculated and calculated The sub-problems of optimal computing frequency and bandwidth allocation optimization are obtained, and the sub-problems of optimal bandwidth allocation and transmission power optimization are calculated, and the optimal transmission power of intelligent terminals and edge servers is calculated. The present invention jointly optimizes multiple variables of wireless communication and model training, and minimizes system energy consumption in the process of federated learning under the limitation of model training performance. Compared with other solutions, the method of the present invention is significantly better than the comparative solution. Can significantly reduce system energy consumption.

As used herein, the terms "comprises," "comprising," or any other variation thereof are intended to cover a non-exclusive inclusion such that a step, method comprising a series of elements includes not only those elements, but also other elements not expressly listed. elements, or also include elements inherent in such steps and methods.

The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be assumed that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deduction or replacement can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (9)

1. a parameter optimization and resource allocation method of a low-energy wireless federated learning system, characterized in that, comprising the following steps: Build a low-energy wireless federated learning system model, including a cloud server, N edge servers and M smart terminals, and establish a federated learning process based on the low-energy wireless federated learning system model; With the optimization goal of minimizing the energy consumption of the system model, limit the number of smart terminal local model training related to the aggregation interval, the calculation frequency of the smart terminal, the system bandwidth allocation, and the transmission power of the smart terminal and the edge server, and establish joint parameter optimization and resource allocation. question, Split the joint parameter optimization and resource allocation problem into aggregation interval optimization sub-problem and calculate the optimal aggregation interval, frequency optimization sub-problem and calculate the optimal computing frequency, bandwidth allocation optimization sub-problem and calculate the optimal bandwidth allocation and transmission The power optimization sub-problem is calculated to obtain the optimal transmission power of the smart terminal and the edge server. 2. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 1, wherein the federated learning process is established based on the low-energy wireless federated learning system model, specifically comprising the following steps: S1. The cloud server broadcasts the global model and its current parameter values to all smart terminals through the edge server; S2. The smart terminal uses local data to iteratively train the received global model. After κ ₁ local iterative training is completed, the smart terminal uploads the updated global model parameters after training, that is, the local model parameters, to the edge server; S3. After all the associated smart terminals upload the parameters, the edge server aggregates the received model parameters into the parameters of the edge model, and distributes them to all the smart terminals associated with the edge server through multicast; S4. After performing κ ₂ iterations on S2 and S3, the edge server uploads the parameters of the edge model to the cloud server; S5. After all the edge servers upload the edge model parameters, the cloud server aggregates the received model parameters, and updates the global model according to the aggregated model parameters; S6. Iteratively execute S1 to S5 until the global model converges or reaches a preset precision value. 3. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 2, characterized in that the method also includes calculating the computing time delay and computing time delay of the smart terminal in local iterative training based on the federated learning process Energy consumption, the rate at which smart terminals upload global model parameters to the edge server, and the multicast rate of the edge server, including: The calculation delay of smart terminal m in a local iterative training is: T _m,cop =v _m CD _m /f _m , where v _m is the number of CPU rounds required by the terminal to process unit bit sample data, and C is the sample data The bit size of D _m is the number of samples of the terminal, f _m is the calculation frequency of the intelligent terminal, and the corresponding calculation energy consumption of the intelligent terminal is expressed as:

Among them, δ is the energy consumption coefficient of the smart terminal; The rate at which smart terminal m uploads local model parameters to edge server n is:

Among them, B _n is the bandwidth of the edge server n, p _m is the transmission power of the intelligent terminal m, g _n,m is the channel gain, N ₀ is the additive noise power; the parameter upload time of the intelligent terminal is T _m,com =W /R _m , where W is the bit size of the uploaded model parameters, and the communication energy consumption of the smart terminal is E _{m, com} = p _m T _{m, com} ; The multicast rate of edge server n is:

Among them, p _n is the transmission power of edge server n, />

is the index set of all intelligent terminals associated with edge server n, the downlink multicast delay of the edge server is T _n,com =W/R _n , and the communication energy consumption of the edge server is E _n,com =p _n T _n,com . 4. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 3, wherein the specific expression of the system model energy consumption is:

5. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 4, wherein the specific expression of the joint parameter optimization and resource allocation problem is:

st(C1)

(C2)

(C3)

(C4)

(C5)

(C6)

(C7)

(C8)

(C9)

(C10)

in,

Represents a collection of aggregation intervals, />

Represents the set of computing frequencies,

Represents a bandwidth collection, />

Indicates the set of transmission power, κ _glob is the number of local model training in one round of global model iteration, f _m,max is the maximum calculation frequency of the terminal, />

and />

is the maximum calculation time and maximum communication time of the intelligent terminal, T _n,max is the maximum delay of the edge server multicast, B _total is the total system bandwidth, B _min and B _max are the minimum bandwidth and maximum bandwidth of the edge server respectively, p _m,max and p _n,max are the maximum transmission power of the intelligent terminal and the maximum transmission power of the edge server, respectively, and E _m,max and E _n,max are the maximum energy consumption of the intelligent terminal and the maximum energy consumption of the edge server, respectively. 6. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 5, wherein the joint parameter optimization and resource allocation problem is split into aggregation interval optimization sub-problems and the optimal aggregation interval is calculated , the specific expression is:

7. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 5, wherein the joint parameter optimization and resource allocation problem is split into frequency optimization sub-problems and calculated to obtain the optimal calculation frequency, The specific expression is:

8. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 5, wherein the joint parameter optimization and resource allocation problem is split into transmission power optimization sub-problems and calculated to obtain intelligent terminal and edge The optimal transmission power of the server, the specific expression is:

9. The parameter optimization and resource allocation method of the low-energy wireless federated learning system according to claim 5, wherein the joint parameter optimization and resource allocation problem is split into bandwidth allocation optimization sub-problems and the optimal bandwidth allocation is calculated , where the optimal bandwidth allocation is achieved using the convex optimization toolbox.

Images