Parameter optimization and resource allocation method for low-energy wireless federated learning system
By constructing a low-energy wireless federated learning system model and jointly optimizing aggregation interval, computing frequency, bandwidth allocation, and transmission power, the problem of limited energy consumption of intelligent terminals in hybrid wireless heterogeneous networks is solved, and the system energy consumption is significantly reduced.
Patent Information
- Application Number
- CN202310196400.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-03
- Publication Date
- 2026-02-13
- Estimated Expiration
- 2043-03-03
AI Technical Summary
In hybrid wireless heterogeneous networks, the energy consumption of smart terminals is limited. How to effectively reduce system energy consumption while ensuring model training performance has become a key issue in federated learning.
A low-energy wireless federated learning system model is constructed. By jointly optimizing the aggregation interval, computation frequency, bandwidth allocation and transmission power, a resource allocation method is designed to minimize system energy consumption. This includes breaking down the problem into sub-problems of aggregation interval optimization, frequency optimization, bandwidth allocation and transmission power optimization, and using a convex optimization toolbox to achieve optimal allocation.
It significantly reduces system energy consumption, outperforming other solutions, and achieves energy consumption optimization under the constraints of model training performance.
Smart Images
Figure CN116418687B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of wireless communication, and particularly relates to a parameter optimization and resource allocation method for a low-energy wireless federated learning system. BACKGROUND
[0002] Compared with a traditional centralized machine learning model training mode, federated learning can make full use of local data and computing resources of distributed intelligent terminals, effectively complete distributed training of a machine learning model under the premise of protecting data privacy of the intelligent terminals and saving communication resources. However, in a hybrid wireless heterogeneous network, the working time of the intelligent terminals is often limited by their limited energy reserves. Therefore, how to effectively reduce system energy consumption under the premise of ensuring model training performance has become one of the problems to be solved by federated learning. SUMMARY
[0003] The application provides a parameter optimization and resource allocation method for a low-energy wireless federated learning system, which is designed to reduce energy consumption of a parameter optimization and resource allocation method, aiming at the fact that in a wireless federated learning system, energy consumption mainly includes computing energy consumption of model training and communication energy consumption of model parameter transmission.
[0004] The technical scheme of the application is as follows:
[0005] A parameter optimization and resource allocation method for a low-energy wireless federated learning system, comprising the following steps:
[0006] A low-energy wireless federated learning system model is constructed, including a cloud server, N edge servers and M intelligent terminals, and a federated learning process is established based on the low-energy wireless federated learning system model;
[0007] A joint parameter optimization and resource allocation problem is established, with the minimum system model energy consumption as an optimization target, and the intelligent terminal local model training frequency related to the aggregation interval, the intelligent terminal computing frequency, the system bandwidth allocation and the transmission power of the intelligent terminal and the edge server being limited.
[0008] The joint parameter optimization and resource allocation problem is divided into an aggregation interval optimization sub-problem, a frequency optimization sub-problem, a bandwidth allocation optimization sub-problem and a transmission power optimization sub-problem, and the optimal aggregation interval, the optimal computing frequency, the optimal bandwidth allocation and the optimal transmission power of the intelligent terminal and the edge server are calculated.
[0009] A further technical scheme of the application is that a federated learning process is established based on a low-energy wireless federated learning system model, specifically comprising the following steps:
[0010] S1, the cloud server broadcasts the global model and its current parameter value to all intelligent terminals through the edge server;
[0011] S2, the intelligent terminal iteratively trains the received global model with local data, and after κ1 times of local iterative training, the intelligent terminal uploads the trained global model parameters, i.e. local model parameters, to the edge server;
[0012] S3, after the associated intelligent terminals upload the parameters, the edge server aggregates the received model parameters into the parameters of the edge model, and distributes them to all intelligent terminals associated with the edge server through multicast;
[0013] S4, after κ2 times of iterative execution of S2 and S3, the edge server uploads the parameters of the edge model to the cloud server;
[0014] S5, after all 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;
[0015] S6, iteratively execute S1 to S5 until the global model converges or reaches a preset precision value.
[0016] Further technical solutions of the application are: the method further comprises calculating the computing time delay and computing energy consumption of the intelligent terminal in local iterative training, the rate of the intelligent terminal uploading the global model parameters to the edge server and the multicast rate of the edge server based on the federated learning process, and specifically comprising:
[0017] The computing time delay of intelligent terminal m in one local iterative training is: m,cop =v m CD m / f m , wherein v m is the number of CPU rounds required by the terminal to process a unit of bit sample data, C is the bit size of the sample data, D m is the number of samples of the terminal, f m is the computing frequency of the intelligent terminal, and the computing energy consumption of the corresponding intelligent terminal is represented as: Where δ is the energy consumption coefficient of the intelligent terminal;
[0018] The rate of intelligent terminal m uploading local model parameters to edge server n is: Where B n is the bandwidth of edge server n, p m is the transmission power of intelligent terminal m, g n,m is the channel gain, and N0 is the additive noise power; the parameter uploading time of the intelligent terminal is T m,com =W / R mwhere W is the bit size of uploading model parameters, and the communication energy consumption of the intelligent terminal is E m,com = p m m,com
[0019] The multicast rate of the edge server n is: where 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, 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 .
[0020] A further technical solution of the present application is that the specific expression of the system model energy consumption is:
[0021]
[0022] A further technical solution of the present application is that the specific expression of the joint parameter optimization and resource allocation problem is:
[0023]
[0024] wherein denotes the aggregation interval set, denotes the calculation frequency set, denotes the bandwidth set, denotes the transmission power set, κ glob is the number of local model training in a round of global model iteration, f m,max is the maximum calculation frequency of the terminal, and are the maximum calculation time and the maximum communication time of the intelligent terminal, T n,max is the maximum delay of the edge server multicast, B total is the total bandwidth of the system, B min and B max are the minimum bandwidth and the maximum bandwidth of the edge server, 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, 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.
[0025] A further technical solution of the present application is that the joint parameter optimization and resource allocation problem is split into an aggregation interval optimization sub-problem, and the optimal aggregation interval is calculated, and the specific expression is:
[0026]
[0027]
[0028] A further technical solution of the present application is to split the joint parameter optimization and resource allocation problem into a frequency optimization sub-problem and calculate the optimal calculation frequency, and the specific expression is:
[0029] A further technical solution of the present application is to split the joint parameter optimization and resource allocation problem into a transmission power optimization sub-problem and calculate the optimal transmission power of the intelligent terminal and the edge server, and the specific expression is:
[0030]
[0031] A further technical solution of the present application is to split the joint parameter optimization and resource allocation problem into a bandwidth allocation optimization sub-problem and calculate the optimal bandwidth allocation, wherein the optimal bandwidth allocation is realized by using a convex optimization toolbox.
[0032] The parameter optimization and resource allocation method of the low-energy wireless federated learning system provided by the present application constructs a low-energy wireless federated learning system model, establishes a federated learning process based on the low-energy wireless federated learning system model, takes minimizing the energy consumption of the system model as the optimization goal, establishes a joint parameter optimization and resource allocation problem, splits the joint parameter optimization and resource allocation problem into an aggregation interval optimization sub-problem and calculates the optimal aggregation interval, a frequency optimization sub-problem and calculates the optimal calculation frequency, a bandwidth allocation optimization sub-problem and calculates the optimal bandwidth allocation, and a transmission power optimization sub-problem and calculates the optimal transmission power of the intelligent terminal and the edge server. The present application jointly optimizes multiple variables of wireless communication and model training, minimizes the system energy consumption in the federated learning process under the limitation of model training performance, and is significantly better than the comparative scheme, which can significantly reduce the system energy consumption. BRIEF DESCRIPTION OF DRAWINGS
[0033] Figure 1 FIG. 1 is a flowchart of the parameter optimization and resource allocation method of the low-energy wireless federated learning system in the embodiment of the present application;
[0034] Figure 2 FIG. 2 is a schematic diagram of the wireless hierarchical federated learning system model structure in the embodiment of the present application;
[0035] Figure 3 FIG. 3 is a comparison diagram of the system energy consumption under different optimization schemes and the number of intelligent terminals associated with the edge server in the embodiment of the present application. DETAILED DESCRIPTION
[0036] The application will be described in further detail below with reference to the drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of the application. In addition, it is to be understood that, for ease of description, only the parts related to the application are shown in the drawings rather than all the structures.
[0037] Before the example embodiments are discussed in more detail, it should be mentioned that some of the example embodiments are described as processes or methods depicted as flow charts. While the steps of the processes are depicted in a sequential order, many of the steps can be performed in parallel, concurrently or at the same time. In addition, the order of the steps can be re-arranged. The processes can be terminated when their operations are completed, but can also have additional steps not included in the figure. The processes can correspond to methods, functions, procedures, subroutines, subprograms, etc.
[0038] As shown in Figure 1 , the parameter optimization and resource allocation method of the low-energy wireless federated learning system in the embodiment includes the following steps:
[0039] A low-energy wireless federated learning system model is constructed, including a cloud server, N edge servers and M intelligent terminals. A federated learning process is established based on the low-energy wireless federated learning system model.
[0040] With the minimization of system model energy consumption as the optimization goal, the number of local model training of intelligent terminals related to aggregation interval, the computing frequency of intelligent terminals, system bandwidth allocation and the transmission power of intelligent terminals and edge servers are limited. A joint parameter optimization and resource allocation problem is established.
[0041] The joint parameter optimization and resource allocation problem is split into an aggregation interval optimization sub-problem and the optimal aggregation interval is calculated, a frequency optimization sub-problem and the optimal computing frequency is calculated, a bandwidth allocation optimization sub-problem and the optimal bandwidth allocation is calculated, and a transmission power optimization sub-problem and the optimal transmission power of intelligent terminals and edge servers is calculated.
[0042] In the specific implementation process, as shown in Figure 2 , the wireless hierarchical federated learning system provided by the embodiment includes a cloud server, N edge servers and M intelligent terminals. The cloud server is connected with each edge server through a high-performance wired backhaul link, and each edge server is connected with each intelligent terminal associated therewith through a wireless communication link.
[0043] Preferably, a federated learning process is established based on the low-energy wireless federated learning system model, specifically including the following steps:
[0044] S1. The cloud server broadcasts the global model and its current parameter values to all smart terminals through the edge server;
[0045] S2. The smart terminal uses local data to iteratively train the received global model. After κ1 local iterations of training, the smart terminal uploads the updated global model parameters, i.e., the local model parameters, to the edge server.
[0046] S3. After all the associated smart terminals have uploaded the parameters, the edge server will aggregate the received model parameters into the parameters of the edge model and distribute them to all the smart terminals associated with the edge server via multicast.
[0047] S4. After iterating S2 and S3 κ2 times, the edge server uploads the parameters of the edge model to the cloud server.
[0048] S5. After all edge model parameters have been uploaded to all edge servers, the cloud server aggregates the received model parameters and updates the global model based on the aggregated model parameters.
[0049] S6. Iterate through S1 to S5 until the global model converges or reaches the preset accuracy value. At this point, the machine learning model training process of this federated learning system is complete.
[0050] In the specific implementation process, the computational latency and energy consumption of the intelligent terminal during local iterative training based on the federated learning process, the rate at which the intelligent terminal uploads global model parameters to the edge server, and the multicast rate of the edge server are specifically included:
[0051] The computation latency of the intelligent terminal m in one local iteration training is: T m,cop =v m CD m / f m , where v m The number of CPU rounds required for the terminal to process a unit bit of sample data, where C is the bit size of the sample data, and D is the number of rounds required. m f is the number of samples from the terminal. m The computing frequency of the smart terminal is represented by the corresponding computing energy consumption of the smart terminal as follows: Where δ is the energy consumption coefficient of the smart terminal;
[0052] The rate at which smart terminal m uploads local model parameters to edge server n is: Among them, B n For the bandwidth of edge server n, p m For the transmission power of the smart terminal m, g n,m Where N is the channel gain, N0 is the additive noise power; the parameter upload time of the smart terminal is T. m,com =W / R mwhere W is the bit size of uploading model parameters, and the communication energy consumption of the intelligent terminal is E m,com = p m T m,com ;
[0053] The multicast rate of the edge server n is: where p n is the transmission power of the edge server 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 , and the communication energy consumption of the edge server is E n,com = p n T n,com .
[0054] In addition, the wireless communication link in the system adopts a "hybrid frequency-time bandwidth allocation" scheme, and the total bandwidth of the system is allocated to each edge server in a frequency division manner, and each intelligent terminal uses the bandwidth allocated by its edge server in a time division manner.
[0055] Further, the specific expression of the system model energy consumption is:
[0056] Embodiments aim to minimize the system energy consumption defined above, and jointly optimize parameters and allocate resources, specifically involving aggregation interval calculation frequency bandwidth allocation and transmission power
[0057] The joint parameter optimization and resource allocation problem can be modeled as:
[0058]
[0059] where denotes the aggregation interval set, denotes the calculation frequency set, denotes the bandwidth set, denotes the transmission power set, κ glob is the number of local model training in a round of global model iteration, f m,max is the maximum calculation frequency of the terminal, and are the maximum calculation time and the maximum communication time of the intelligent terminal, T n,max is the maximum delay of the edge server multicast, B total is the total bandwidth of the system, B min and B max are the minimum bandwidth and the maximum bandwidth of the edge server, respectively, and pm,max and p n,max are the maximum transmission power of the intelligent terminal and the maximum transmission power of the edge server, respectively, 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.
[0060] Preferably, the embodiment splits the above joint optimization problem into four sub-problems of aggregated interval optimization, computing frequency optimization, bandwidth allocation optimization and transmission power optimization:
[0061] The joint parameter optimization and resource allocation problem is split into an aggregated interval optimization sub-problem and the optimal aggregated interval is calculated, and the specific expression is:
[0062]
[0063]
[0064] The joint parameter optimization and resource allocation problem is split into a frequency optimization sub-problem and the optimal computing frequency is calculated, and the specific expression is:
[0065] The joint parameter optimization and resource allocation problem is split into a transmission power optimization sub-problem and the optimal transmission power of the intelligent terminal and the edge server is calculated, and the specific expression is:
[0066]
[0067] In addition, the joint parameter optimization and resource allocation problem is split into a bandwidth allocation optimization sub-problem and the optimal bandwidth allocation is calculated, wherein the optimal bandwidth allocation is realized by using a convex optimization toolbox.
[0068] The flow of the parameter optimization and resource allocation method based on low-energy wireless federated learning is summarized in Algorithm 1 as follows:
[0069]
[0070]
[0071] In order to better reflect the effectiveness of the present application, simulation experiments are carried out. Figure 3 The trend of system energy consumption with the number of intelligent terminals associated with the edge server under different optimization schemes is shown. The total number of system intelligent terminals M is set to 2000 and 3000 respectively in the experiment. It can be observed that the parameter optimization and resource allocation method proposed in the present application can be consistent with the exhaustive search algorithm of the joint optimization problem, and is significantly better than the comparative scheme, and can significantly reduce the system energy consumption.
[0072] As can be seen from the embodiments, the parameter optimization and resource allocation method of the low-energy wireless federated learning system provided by the application constructs a low-energy wireless federated learning system model, establishes a federated learning process based on the low-energy wireless federated learning system model, takes minimizing the energy consumption of the system model as an optimization goal, establishes a joint parameter optimization and resource allocation problem, splits the joint parameter optimization and resource allocation problem into an aggregation interval optimization sub-problem and calculates an optimal aggregation interval, a frequency optimization sub-problem and calculates an optimal calculation frequency, a bandwidth allocation optimization sub-problem and calculates an optimal bandwidth allocation, and a transmission power optimization sub-problem and calculates optimal transmission power of the intelligent terminal and the edge server. The application jointly optimizes multiple variables of wireless communication and model training, minimizes the system energy consumption in the federated learning process under the limitation of model training performance, and is significantly better than the comparative scheme, which can significantly reduce the system energy consumption.
[0073] In this document, the terms "comprise", "comprising", or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article of manufacture, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article of manufacture, or apparatus.
[0074] The above is a further detailed description of the application in combination with specific preferred embodiments, and cannot be regarded as limiting the specific implementation of the application to these descriptions. For ordinary skilled persons in the art to which the application belongs, a number of simple deductions or replacements can be made without departing from the concept of the application, and all of them should be regarded as falling within the protection scope of the application.
Claims
1. A method for parameter optimization and resource allocation of a low-energy wireless federated learning system, characterized in that, The method comprises the following steps: A low-energy wireless federated learning system model is constructed, including a cloud server, N edge servers and M intelligent terminals, and a federated learning process is established based on the low-energy wireless federated learning system model; A joint parameter optimization and resource allocation problem is established, taking the minimization of the energy consumption of the system model as an optimization objective, limiting the number of local model training of the intelligent terminal related to the aggregation interval, the computing frequency of the intelligent terminal, the system bandwidth allocation and the transmission power of the intelligent terminal and the edge server, The joint parameter optimization and resource allocation problem is split into an aggregation interval optimization sub-problem and the optimal aggregation interval is calculated, a frequency optimization sub-problem and the optimal computing frequency is calculated, a bandwidth allocation optimization sub-problem and the optimal bandwidth allocation is calculated and a transmission power optimization sub-problem and the optimal transmission power of the intelligent terminal and the edge server is calculated; The federated learning process is established based on the low-energy wireless federated learning system model, and specifically comprises the following steps: S1. The cloud server broadcasts a global model and its current parameter value to all intelligent terminals through the edge server; S2. The intelligent terminal iteratively trains the received global model using local data, and after κ1 local iteration training, the intelligent terminal uploads the trained global model parameters, i.e., local model parameters, to the edge server; S3. After the associated intelligent terminals all upload the parameters, the edge server aggregates the received model parameters into the parameters of an edge model and distributes them to all intelligent terminals associated with the edge server through multicasting; S4. After κ2 iterations of S2 and S3, the edge server uploads the parameters of the edge model to the cloud server; S5. After all 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. S1 to S5 are iteratively executed until the global model converges or reaches a preset accuracy value; The method further comprises calculating the computing delay and computing energy consumption of the intelligent terminal in local iteration training, the rate of the intelligent terminal uploading global model parameters to the edge server and the multicasting rate of the edge server based on the federated learning process, and specifically comprises: The calculation time delay of the intelligent terminal m in one local iteration training is T m,cop = v m CD m / f m , wherein v m is the CPU round required for the terminal to process one bit sample data, C is the bit size of the sample data, D m is the sample quantity of the terminal, f m is the calculation frequency of the intelligent terminal, and the calculation energy consumption of the corresponding intelligent terminal is represented as: wherein δ is the energy consumption coefficient of the intelligent terminal. The rate of the intelligent terminal m uploading the local model parameters to the edge server n is: Wherein, 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, and N0 is the additive noise power; the parameter uploading time of the intelligent terminal is T m,com =W / R m , wherein W is the bit size of the uploaded model parameters, and the communication energy consumption of the intelligent terminal is E m,com =p m T m,com ; The multicast rate of the edge server n is: wherein p n is the transmission power of the edge server n, is the index set of all intelligent terminals associated with the 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 ; The specific expression of the system model energy consumption is: The specific expression of the joint parameter optimization and resource allocation problem is: wherein, denotes a set of aggregation intervals, denotes a set of computing frequencies, denotes a set of bandwidths, denotes a set of transmission powers, κ glob is the number of local model training times in a round of global model iteration, f m,max is the maximum computing frequency of the terminal, and are the maximum computing time and the maximum communication time of the intelligent terminal, T n,max is the maximum delay of the edge server groupcast, B total is the total bandwidth of the system, B min and B max are the minimum bandwidth and the maximum bandwidth of the edge server, 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, 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; The joint parameter optimization and resource allocation problem is split into an aggregation interval optimization sub-problem and the optimal aggregation interval is calculated, and the specific expression is: The joint parameter optimization and resource allocation problem is split into a frequency optimization subproblem and the optimal computation frequency is computed, expressed as: The joint parameter optimization and resource allocation problem is split into a transmission power optimization sub-problem and the optimal transmission power of the intelligent terminal and the edge server is calculated, and the specific expression is: The joint parameter optimization and resource allocation problem is split into a bandwidth allocation optimization sub-problem and the optimal bandwidth allocation is calculated, and the optimal bandwidth allocation is realized by using a convex optimization toolbox.