Federal learning channel resource optimization method and system based on information age

Abstract

The invention relates to a federated learning channel resource optimization method and system based on information age, and the method comprises the steps: obtaining system parameters and environment parameters of a federated learning network, and constructing a federated learning system model; defining an information age for the equipment and establishing an information age updating mechanism; according to the federated learning system model, by taking total information age minimization as a target, establishing an optimization problem including an energy consumption constraint and a sub-channel allocation constraint; decomposing the optimization problem into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem; solving the resource allocation sub-problem to obtain an optimal power allocation scheme; and solving the sub-problems of equipment selection and sub-channel allocation to obtain an optimal pairing scheme of the equipment and the sub-channels. The method effectively improves the convergence speed and precision of the model, guarantees the fairness of equipment selection, and is suitable for a federated learning system in a non-independent identically distributed data environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of federated learning and wireless communication network intersection, in particular to a federated learning channel resource optimization method and system based on information age. BACKGROUND

[0002] With the rapid development of Internet of Things and edge computing, federated learning as a distributed machine learning paradigm can collaboratively train global models using local data from edge devices while protecting user privacy. However, in a wireless network environment, the performance of federated learning is restricted by multiple factors such as limited communication resources, device heterogeneity, and non-identically distributed data distribution.

[0003] In existing research, some works focus on delay minimization, device selection, and resource allocation in federated learning. For example, some solutions optimize communication and computing resource allocation to reduce single-round training time. Other research proposes device selection strategies based on contribution or channel state. However, these methods have significant limitations. First, if only delay minimization is targeted, devices with good channel conditions may be repeatedly selected, causing model bias or overfitting. Second, in a non-identically distributed data environment, random or channel-driven device selection strategies can introduce gradient bias, affecting model convergence speed and accuracy.

[0004] In summary, existing methods fail to effectively balance the relationship between delay and model convergence, lacking a comprehensive index that reflects both time efficiency and fair device selection. Therefore, there is an urgent need for a joint optimization scheme that considers delay, energy consumption, device fairness, and model convergence performance in a non-identically distributed data environment. SUMMARY

[0005] The purpose of the present application is to provide a federated learning channel resource optimization method and system based on information age, which introduces information age as an indicator of device data freshness, jointly optimizes device selection and subchannel allocation to improve model convergence speed and accuracy while ensuring fairness among devices.

[0006] To achieve the above purpose, the present application provides the following solutions:

[0007] A federated learning channel resource optimization method based on information age, comprising:

[0008] Obtain system parameters and environmental parameters of the federated learning network, and construct a federated learning system model, wherein the federated learning system model includes a central server, devices, and subchannels;

[0009] defining an information age of the device and establishing an information age updating mechanism, wherein the updating mechanism is that if the device is selected in the current round, the information age is reset to zero, and if the device is not selected, the information age is accumulated by incrementing the total time consumption in the current round;

[0010] establishing, according to the federated learning system model, an optimization problem containing energy consumption constraints and sub-channel allocation constraints with the goal of minimizing the total information age;

[0011] decomposing the optimization problem into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem;

[0012] solving the resource allocation sub-problem to obtain an optimal power allocation scheme;

[0013] solving the device selection and sub-channel allocation sub-problem to obtain an optimal pairing scheme of the devices and the sub-channels.

[0014] Optionally, the information age updating mechanism is:

[0015] ;

[0016] wherein, is the information age of the device n in the tth round, is the information age of the device n in the (t-1)th round, is a sub-channel allocation indicator variable in the tth round, is the total time consumption in the tth round, and K is the number of sub-channels.

[0017] Optionally, establishing the optimization problem containing energy consumption constraints and sub-channel allocation constraints with the goal of minimizing the total information age includes:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] wherein, is the energy consumption of the device, is a maximum allowed value, and N is the number of devices.

[0023] Optionally, decomposing the optimization problem into the resource allocation sub-problem and the device selection and sub-channel allocation sub-problem includes that the resource allocation sub-problem optimizes power allocation coefficients to minimize the single-round time consumption given the device selection and the sub-channel allocation;

[0024] The device selection and subchannel allocation sub-problem selects devices and allocates subchannels by a matching algorithm, and optimizes the total information age.

[0025] Optionally, solving the resource allocation sub-problem to obtain the optimal power allocation scheme comprises: taking the minimum time consumption as an objective function and solving according to the KKT condition to obtain the optimal power allocation coefficient, and obtaining the optimal power allocation scheme through the optimal power allocation coefficient.

[0026] The objective function of the minimum time consumption is:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] wherein, represents the total time consumed by device n being allocated to subchannel k, represents the computing time of the device, is the communication time, is the number of CPU cycles required per sample, is the number of samples of device n, is the available CPU cycle number of device n, D is the model size, and B is the subchannel bandwidth, is the power allocation coefficient, is the maximum transmission power of device n, is the channel gain.

[0032] Optionally, solving the device selection and subchannel allocation sub-problem to obtain the optimal pairing scheme of the device and the subchannel comprises:

[0033] A bipartite graph model is constructed, wherein the left node set is the device, the right node set is the union of physical subchannels and virtual subchannels, and the edge is the connection between the device and the physical subchannel and the virtual subchannel.

[0034] The maximum weight matching of the bipartite graph is solved according to the Hungarian algorithm to obtain the optimal pairing scheme of the device and the subchannel.

[0035] Optionally, the connection weight between the device and the physical subchannel is:

[0036] ;

[0037] The connection weight between the device and the virtual subchannel is:

[0038] ;

[0039] wherein, is the device n information age of the t-1th round, denotes the total time consumed by device n being assigned to subchannel k, is the energy consumption of the device, is the maximum allowed value, is the adjustable weight coefficient.

[0040] The application also provides a federated learning channel resource optimization system based on information age, comprising:

[0041] a parameter acquisition module, configured to acquire system parameters and environment parameters of a federated learning network, construct a federated learning system model, and construct the federated learning system model according to the system parameters, wherein the federated learning system model comprises a central server, a device and a subchannel;

[0042] an information age modeling module, configured to define information age for the device and establish an information age updating mechanism, wherein if the device is selected in the current round, the information age is reset to zero, and if the device is not selected, the information age is accumulated by incrementing the total time consumption in the current round;

[0043] a problem construction module, configured to establish an optimization problem comprising energy consumption constraints and subchannel allocation constraints according to the federated learning system model, with the goal of minimizing total information age;

[0044] a problem decomposition module, configured to decompose the optimization problem into a resource allocation subproblem and a device selection and subchannel allocation subproblem;

[0045] a resource allocation solving module, configured to solve the resource allocation subproblem and acquire an optimal power allocation scheme;

[0046] a matching scheduling module, configured to solve the device selection and subchannel allocation subproblem and acquire an optimal pairing scheme of the device and the subchannel.

[0047] The beneficial effects of the present application are: the present application introduces information age as a quantitative indicator of device data freshness, effectively balances the delay optimization and model convergence performance in the federated learning process. By constructing a system total information age minimization problem, the complex joint optimization problem is decomposed into two sub-problems of resource allocation and device selection for solving, which significantly improves the optimization efficiency. In terms of resource allocation, the present application uses KKT condition to derive the closed-form solution of power allocation, avoids the high computational complexity of traditional iterative algorithm, can quickly obtain the optimal power allocation scheme, and effectively reduces the time consumption of single round communication. In terms of device selection and sub-channel allocation, the maximum weight matching algorithm is innovatively used, a bipartite graph model containing physical sub-channels and virtual sub-channels is constructed, and multi-dimensional factors such as device information age, time efficiency and energy consumption are comprehensively considered, so as to realize the optimal matching of devices and channels. This method not only ensures the fairness of device selection, avoids the model deviation problem caused by some devices being selected for a long time, but also dynamically adapts to the change of network environment. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0049] Figure 1 The flow chart of the federated learning channel resource optimization method based on information age according to an embodiment of the present application;

[0050] Figure 2 The federated learning system model diagram according to an embodiment of the present application;

[0051] Figure 3 The extended bipartite graph for solving the device selection and sub-channel allocation problem according to an embodiment of the present application;

[0052] Figure 4 The architecture diagram of the federated learning channel resource optimization system based on information age according to an embodiment of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0055] The embodiment provides a federated learning channel resource optimization method based on information age, which comprises the following steps:

[0056] System parameters and environment parameters of a federated learning network are acquired, and a federated learning system model is constructed, wherein the federated learning system model comprises a central server, devices and sub-channels;

[0057] Information age of the devices is defined and an information age updating mechanism is established, wherein if a device is selected in the current round, the information age is reset to zero, and if the device is not selected, the information age is accumulated by incrementing the total time consumption in the current round;

[0058] According to the federated learning system model, an optimization problem containing energy consumption constraints and sub-channel allocation constraints is established with the minimum total information age as the target;

[0059] The optimization problem is decomposed into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem;

[0060] The resource allocation sub-problem is solved to obtain an optimal power allocation scheme;

[0061] The device selection and sub-channel allocation sub-problem is solved to obtain an optimal pairing scheme of the devices and the sub-channels.

[0062] Further, the federated learning system model comprises:

[0063] A device set is defined as N={1, 2, …, N}, and a sub-channel set is defined as K={1, 2, …, K}, wherein N≥K;

[0064] Each device holds a local non-independent and identically distributed data set, and the server broadcasts a global model to all devices in each round of communication;

[0065] According to the condition that the number of sub-channels is limited, only part of the devices are selected to participate in model aggregation in each round.

[0066] Further, information age of each device is defined , which represents the freshness of the data of the device;

[0067] According to the information age updating rule, if a device is selected in the current round, the information age of the device is reset to zero, and if the device is not selected, the information age is accumulated by incrementing the total time consumption in the current round.

[0068] The information age updating mechanism is:

[0069] ;

[0070] wherein, is the information age of device n in the tth round, is the information age of device n in the (t−1)th round, is the subchannel assignment indicator variable in the tth round, is the total time consumption in the tth round, K is the number of subchannels.

[0071] Further, an optimization problem is established to minimize the total information age with energy consumption constraint and subchannel assignment constraint, including:

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] wherein, is the energy consumption of device n, is the maximum allowed value, N is the number of devices.

[0077] Further, the resource allocation subproblem optimizes the power allocation coefficient to minimize the single round time consumption given the device selection and subchannel assignment;

[0078] The device selection and subchannel assignment subproblem optimizes the total information age by selecting devices and assigning subchannels through a matching algorithm.

[0079] Further, the resource allocation subproblem is solved to obtain the optimal power allocation scheme, including: the resource allocation subproblem is specified as a power allocation problem, and a corresponding convex optimization model is established, the objective function is to minimize the time consumption, and the optimal power allocation coefficient is obtained by solving according to the KKT condition, and the optimal power allocation scheme is obtained through the optimal power allocation coefficient;

[0080] The objective function for minimizing the time consumption is:

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] wherein, represents the total time consumed by device n being assigned to subchannel k, represents the computation time of the device, which is usually dependent on the local data amount and computation capability, and is independent of the power allocation, represents the communication time, which is dependent on the transmission rate, and the rate is related to the allocated power, represents the number of CPU cycles required for each sample, represents the number of samples of the device n, represents the number of available CPU cycles of the device n, D is the model size, and B is the subchannel bandwidth, is a power allocation coefficient, is the maximum transmission power of the device n, is the channel gain.

[0086] The feasibility of the solution is verified in the embodiment, and it is ensured that the energy constraint is satisfied, wherein the energy constraint expression is:

[0087] ;

[0088] wherein, is the maximum allowed energy consumption of the device.

[0089] Further, the device selection and subchannel allocation sub-problem are solved to obtain the optimal pairing scheme of the device and the subchannel, including:

[0090] A bipartite graph model is constructed, wherein the left node set is the device, the right node set is the union set of the physical subchannel and the virtual subchannel, and the edge is the connection between the device and the physical subchannel and the virtual subchannel;

[0091] The maximum weight matching of the bipartite graph is solved according to the Hungarian algorithm to obtain the optimal pairing scheme of the device and the subchannel.

[0092] Further, the connection weight between the device and the physical subchannel is:

[0093] ;

[0094] The connection weight between the device and the virtual subchannel is:

[0095] ;

[0096] wherein, is the information age of the device n in the t-1th round, represents the total time consumed by the device n when being allocated to the subchannel k, is the energy consumption of the device, is the maximum allowed value, is an adjustable weight coefficient.

[0097] The method of the embodiment is further described below with reference to the accompanying drawings:

[0098] As Figure 1 shown, a federated learning channel resource optimization method based on information age includes:

[0099] S1: Obtain system parameters and environmental parameters of the federated learning network, and construct a federated learning system model containing a central server, multiple terminal devices, and a limited number of sub-channels, wherein the terminal devices hold non-identically distributed data sets;

[0100] S2: Define an information age index for each device, which increases over time when the device is not selected to participate in model aggregation, and is reset to zero when selected;

[0101] S3: Establish an optimization problem containing energy consumption constraints and sub-channel allocation constraints with the goal of minimizing the total system information age;

[0102] S4: Decompose the optimization problem into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem. For the resource allocation sub-problem, use the KKT condition to solve the optimal power allocation scheme;

[0103] S5: For the device selection and sub-channel allocation sub-problem, use the maximum weight matching algorithm to achieve the optimal pairing of devices and sub-channels.

[0104] Further, obtain the system parameters of the federated learning network, including device computing capacity, communication bandwidth, energy constraints, and non-identically distributed data characteristic parameters of each device. Among them, the total number of devices N, the total number of sub-channels K, the device computing capacity , the device local data volume , the model size D, the sub-channel bandwidth B, and the maximum transmission power of the device, where .

[0105] The environmental parameters include: the channel gain between each device and the server, the device energy consumption coefficient , and the number of CPU cycles required per sample .

[0106] Further, constructing a federated learning system model includes:

[0107] Consider a federated learning network system, as shown in Figure 2 . The system contains a central server and N terminal devices, and the device set is defined as ;

[0108] There are K sub-channels in the system, and the sub-channel set is defined as , where .

[0109] Each device holds a local non-iid dataset, and the server broadcasts the global model to all devices in each round of communication.

[0110] The federated learning process proceeds through successive rounds of communication . Each round contains the following phases:

[0111] S1: The server broadcasts the current global model to all devices;

[0112] S2: Each device trains on its local dataset using the received model, computes a local model update ;

[0113] S3: The server collects local model updates from a subset of devices, updates the global model via federated averaging:

[0114] ;

[0115] where is the subset of devices selected to participate in aggregation in round . According to the limited number of sub-channels condition, only a subset of devices is selected to participate in model aggregation in each round.

[0116] Further, the local loss function is defined as:

[0117] ;

[0118] where is the loss function, is the th sample of device .

[0119] Further, based on the network parameters and channel state, device selection calculation and information age update are constructed, and the model is assembled, which specifically includes:

[0120] Based on the current channel gain and the sub-channel bandwidth B, the device communication rate is calculated;

[0121] Based on the device communication rate and the model size D, the communication time is constructed;

[0122] Based on the device computing power and the local data volume, the calculation time is constructed;

[0123] Based on the device selection decision , the total time consumption , the information age update model is constructed;

[0124] Based on the age of each device information Construct the system total information age optimization target.

[0125] Further, the device model transmission demand in the federal learning network communication scenario, the communication time of a single device n According to the Shannon capacity formula:

[0126] ;

[0127] Among them, The communication time of device n on subchannel k is expressed in seconds, which is determined by the Shannon formula and reflects the time required for device to complete model transmission under the current channel condition and power allocation.

[0128] Further, based on the calculation time demand of the device local calculation characteristics, the calculation time of device n is determined by the local data volume and the calculation ability, and the expression is:

[0129] ;

[0130] Among them, The calculation time of device n is expressed in seconds, which ensures that the device can complete the calculation task of local model training.

[0131] Further, the dynamic change demand of device data freshness in federal learning, the information age (AoI) update rule of a single device n is:

[0132] ;

[0133] Among them, The information age of device n in the tth round is expressed in time, which specifies two update conditions: when the device is selected to participate in aggregation, the information age is reset to zero; Otherwise, the information age increases by the total time consumption of this round.

[0134] Further, the system total information age optimization target is to minimize the sum of the information ages of all devices, and the expression is as follows:

[0135] ;

[0136] Among them, The instantaneous information age of the nth device in the tth round is expressed in time; N is the total number of devices in the federal learning system. This optimization target ensures that the data freshness of all devices in the system is as high as possible, thereby improving the model convergence performance of the entire federal learning system.

[0137] Further, the current information age optimization method based on federal learning, the problem is defined as follows:

[0138] ;

[0139] ;

[0140] ;

[0141] ;

[0142] where K denotes the total number of sub-channels in the federated learning network, is the sub-channel allocation decision variable, t represents the communication round, specifies that at most one device can be allocated on any single sub-channel k to avoid inter-device channel collision.

[0143] Thus, an optimization problem is established to minimize the total information age of the system, subject to energy consumption constraints and sub-channel allocation constraints.

[0144] Further, the complex optimization problem is decomposed into two relatively independent sub-problems, including:

[0145] The resource allocation sub-problem optimizes the power allocation coefficient to minimize the time consumption in a single round, given the device selection and sub-channel allocation;

[0146] The device selection and sub-channel allocation sub-problem selects devices and allocates sub-channels through a matching algorithm to indirectly optimize the total information age.

[0147] Further, solving the resource allocation sub-problem according to the KKT condition includes:

[0148] By introducing an intermediate variable and using the formula to transform, the resource allocation sub-problem, i.e., the power allocation problem, can be modeled as a convex optimization problem. Since this problem satisfies the Slater condition, the KKT condition is applied to derive the closed-form optimal solution. The objective function is to minimize the time consumption, where the objective function expression is:

[0149] ;

[0150] where represents the total time consumed by device n allocated to sub-channel k, and the calculation formula is:

[0151] ;

[0152] where, represents the computation time of the device, which is usually dependent on the local data volume and computing power, and is independent of power allocation. While the communication time depends on the transmission rate, which is related to the allocated power.

[0153] In the above application KKT condition solving process, construct the following Lagrange function and derivation:

[0154] ;

[0155] Through the above variable substitution, the Lagrange function is transformed into the form about , and the partial derivative of is taken and set to zero, to obtain the following transcendental equation:

[0156] ;

[0157] The above transcendental equation cannot be directly solved by conventional algebraic methods, and the Lambert W function needs to be used to obtain the optimal power allocation coefficient ;

[0158] Verify the feasibility of the solution to ensure that the energy constraint is met, that is, the total energy consumption of the device at each communication round must satisfy . The total energy is composed of the calculation energy consumption and the communication energy consumption :

[0159] ;

[0160] ;

[0161] Among them, is the power consumption coefficient, is the power allocation coefficient, is the communication time.

[0162] In summary, the energy constraint expression is:

[0163] ;

[0164] Among them, is the maximum allowed energy consumption of the device.

[0165] Further, for the device selection and subchannel allocation subproblem, an exact algorithm based on the maximum weight bipartite graph matching is adopted to realize the multi-objective optimal pairing between the device and the subchannel considering the information age, delay and energy efficiency. This method converts the combinatorial optimization problem into a standard graph theory problem, ensuring that the global optimal solution is obtained in polynomial time. The specific implementation steps are as follows:

[0166] S1: Construct an extended bipartite graph model;

[0167] S2: Design a comprehensive weight function;

[0168] ​​S3: Apply the Hungarian algorithm to solve the maximum weight matching.

[0169] Further, to achieve optimal pairing, first model the device selection and subchannel allocation problem as a complete bipartite graph , as shown in the accompanying Figure 3 . The left node set N represents all devices; to handle the constraint that the number of devices is greater than the number of physical subchannels , introduce a virtual subchannel, denoted as the set . Thus, the right node set , where . is the set of edges, specifically defined as:

[0170] ;

[0171] A key practical constraint is that if a device is matched to a virtual subchannel, it means that it is not selected in the current round and does not participate in aggregation.

[0172] Further, to guide the matching result to simultaneously optimize information age, latency, and energy consumption, define a comprehensive weight for each edge in the bipartite graph . The design of this function is the core of the method, as it quantifies the immediate benefits of different pairing decisions.

[0173] For a physical subchannel , the weight expression is:

[0174] ;

[0175] where is the device 's information age from the previous round, and devices with high AoI are preferred to improve data freshness. is the optimal completion time of device on subchannel , which is solved by the resource allocation subproblem in S5, and its reciprocal is used to reward low-latency transmissions. is the corresponding optimal energy consumption, which is used to reward devices with high energy efficiency (more remaining energy). is a tunable non-negative weight coefficient used to balance the relative importance of the three optimization objectives.

[0176] For a virtual subchannel , the weight expression is:

[0177] ;

[0178] This weight is negative and only related to AoI. This design implements an "action mask" mechanism: it punishes the decision of leaving high-AoI devices idle, thus guiding the algorithm to preferentially activate these devices. At the same time, it provides a controlled "inaction" option for the system, ensuring that the algorithm can still produce a legal solution when resources are extremely limited.

[0179] The maximum weight matching of the above bipartite graph model is the globally optimal device-subchannel allocation scheme. The classic Hungarian Algorithm is used to solve it, which can find the exact optimal solution within time complexity, combining theoretical guarantee and engineering feasibility.

[0180] The solving steps are as follows:

[0181] S1: Construct the weight matrix: initialize a matrix . Among them, the first column corresponds to the physical subchannel, and the element ; the latter column corresponds to the virtual subchannel, and the element .

[0182] S2: Algorithm initialization: initialize the top label array, left top label array , and right top label array . Among them, the top label is used to construct the equivalent subgraph and ensure the optimality condition. Initialize the matching array, right node matching array , indicating that all are not matched at the beginning.

[0183] S3: Iterative search for perfect matching: find the matching right subchannel node for each left device node . This process finds an augmented path in the equivalent subgraph through breadth-first search. Key operations include:

[0184] S31: Dynamically adjust the top label to expand the range of equivalent subgraphs, so as to find new matching possibilities;

[0185] S32: When an unmatched right node is found, update the matching relationship by backtracking the recorded path information (usually stored in array).

[0186] S4: Output the final allocation scheme: for each right node , if , then device is allocated to subchannel ; for right node , if , then device Not selected (assigned to a virtual subchannel).

[0187] Through the above steps, the optimal matching scheme of the device and the subchannel can be obtained, so that the total weight is maximized, thereby minimizing the total information age of the system.

[0188] The present embodiment can use the Hungarian algorithm to optimize the device selection and channel allocation strategy by modeling the device selection and subchannel allocation problem as a bipartite graph matching problem, so as to dynamically adapt to the network state changes and minimize the total information age of the system.

[0189] In a federated learning network environment, as the number of devices and subchannels increases, the complexity of the matching problem grows combinatorially. Moreover, the network environment is dynamically changing, and under the condition of heterogeneous device data distribution and channel condition changes, it is difficult to solve the optimal scheduling strategy using traditional optimization methods. To address this problem, a scheduling method based on maximum weight matching is proposed. The core of the method is based on the Hungarian algorithm, which constructs a bipartite graph model containing physical subchannels and virtual subchannels, and comprehensively considers multi-dimensional factors such as device information age, time efficiency and energy consumption, to achieve the optimal matching of devices and channels.

[0190] As shown in Figure 4 The present embodiment also provides a federated learning channel resource optimization system based on information age, which comprises:

[0191] A parameter acquisition module is configured to acquire system parameters and environment parameters of a federated learning network, and to construct a federated learning system model, wherein the federated learning system model comprises a central server, devices and subchannels.

[0192] An information age modeling module is configured to define information age for devices and to establish an information age updating mechanism, wherein if a device is selected in the current round, the information age is reset to zero, and if the device is not selected, the information age is accumulated by incrementing the total time consumption in the current round.

[0193] A problem construction module is configured to establish an optimization problem containing energy consumption constraints and subchannel allocation constraints based on the federated learning system model, with the goal of minimizing the total information age.

[0194] A problem decomposition module is configured to decompose the optimization problem into a resource allocation subproblem and a device selection and subchannel allocation subproblem.

[0195] A resource allocation solving module is configured to solve the resource allocation subproblem and obtain an optimal power allocation scheme.

[0196] A matching scheduling module is configured to solve the device selection and subchannel allocation subproblem and obtain an optimal pairing scheme of devices and subchannels.

[0197] Specifically, the system modeling module constructs a federated learning framework containing a single server and multiple terminal devices, configures non-independent and identically distributed data sets, and defines a complete process of model broadcasting, local training and model aggregation; the age modeling module establishes an information age dynamic updating rule based on device selection state and communication time delay, and integrates a maximum energy consumption constraint; the problem construction and decomposition module decouples the original mixed integer nonlinear programming problem into two sub-problems which can be independently solved; the resource allocation solving module converts the resource allocation problem into a convex optimization problem through variable substitution, and derives a closed-form optimal solution by applying KKT conditions and Lambert W function; the matching scheduling module constructs a complete bipartite graph model, introduces a constraint that the number of devices is greater than the number of sub-channels, designs a comprehensive weight function to balance information age, time delay and energy efficiency, adopts the Hungarian algorithm to solve the maximum weight matching, and ensures that a globally optimal device-subchannel pairing scheme is obtained within a polynomial time.

[0198] The above-described embodiments are merely descriptions of the preferred modes of the present application, and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.

Claims

1. A method for optimizing channel resources of federated learning based on information age, characterized in that, The method comprises the following steps: acquiring system parameters and environment parameters of a federated learning network, and constructing a federated learning system model, wherein the federated learning system model comprises a central server, devices and sub-channels; defining an information age for the devices and establishing an information age updating mechanism, wherein if the devices are selected in the current round, the information age is reset to zero, and if the devices are not selected, the information age is accumulated by incrementing the total time consumption in the current round; establishing an optimization problem comprising energy consumption constraints and sub-channel allocation constraints, aiming at minimizing the total information age according to the federated learning system model; decomposing the optimization problem into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem; solving the resource allocation sub-problem to obtain an optimal power allocation scheme; solving the device selection and sub-channel allocation sub-problem to obtain an optimal pairing scheme of the devices and the sub-channels.

2. The information age based federated learning channel resource optimization method of claim 1, wherein, The information age updating mechanism comprises: ； wherein, is the age of device n information for the tth round, is the age of device n information for the t-1th round, is the subchannel assignment indicator variable for the tth round, is the total time consumption for the tth round, K is the number of subchannels, k is the subchannel k.

3. The method of claim 2, wherein, establishing an optimization problem comprising energy consumption constraints and sub-channel allocation constraints, aiming at minimizing the total information age comprises: ； ； ； ； wherein, is the energy consumption of the device, is the maximum allowed value, N is the number of devices.

4. The method of claim 1, wherein, decomposing the optimization problem into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem comprises: the resource allocation sub-problem optimizes power allocation coefficients to minimize the time consumption in a single round under given device selection and sub-channel allocation; the device selection and sub-channel allocation sub-problem optimizes the total information age by selecting devices and allocating sub-channels through a matching algorithm.

5. The method of claim 1, wherein, Solving the resource allocation sub-problem to obtain an optimal power allocation scheme comprises: taking minimizing time consumption as an objective function and solving according to KKT conditions to obtain optimal power allocation coefficients, and obtaining the optimal power allocation scheme through the optimal power allocation coefficients; the objective function of minimizing time consumption is: ； ； ； ； where, denotes the total time consumed by device n being allocated to sub-channel k, denotes the computation time of the device, is the communication time, is the number of CPU cycles required per sample, is the number of samples of device n, is the number of available CPU cycles of device n, D is the model size, and B is the sub-channel bandwidth, is the power allocation factor, is the maximum transmit power of device n, is the channel gain.

6. The information age based federated learning channel resource optimization method of claim 1, wherein, Solving the device selection and sub-channel allocation sub-problem to obtain an optimal pairing scheme of the devices and the sub-channels comprises: constructing a bipartite graph model, wherein the left node set is the devices, the right node set is the union of physical sub-channels and virtual sub-channels, and the edges are the connections between the devices and the physical sub-channels and the virtual sub-channels; solving the maximum weight matching of the bipartite graph according to the Hungarian algorithm to obtain the optimal pairing scheme of the devices and the sub-channels.

7. The method of claim 6, wherein, The connection weight between the devices and the physical sub-channels is: ； The connection weight between the devices and the virtual sub-channels is: ； wherein, is the age of device n information for the t-1th round, denotes the total time consumed for device n to be assigned to sub-channel k, is the energy consumption of the device, is the maximum allowed value, is an adjustable weight coefficient.

8. A system for implementing the method for information age based federated learning channel resource optimization according to any one of claims 1-7, wherein The method comprises the following steps: a parameter acquisition module is configured to acquire system parameters and environment parameters of a federated learning network, and construct a federated learning system model, wherein the federated learning system model comprises a central server, devices and sub-channels; an information age modeling module is configured to define an information age for the devices and establish an information age updating mechanism, wherein if the devices are selected in the current round, the information age is reset to zero, and if the devices are not selected, the information age is accumulated by incrementing the total time consumption in the current round; a problem construction module is configured to establish an optimization problem comprising energy consumption constraints and sub-channel allocation constraints, aiming at minimizing the total information age according to the federated learning system model; a problem decomposition module configured to decompose the optimization problem into a resource allocation sub-problem and a device selection and sub-channel allocation sub-problem; a resource allocation solution module configured to solve the resource allocation sub-problem to obtain an optimal power allocation scheme; a matching scheduling module configured to solve the device selection and sub-channel allocation sub-problem to obtain an optimal pairing scheme of the devices and the sub-channels.

Images