Search

CN-116471612-B - Method for jointly optimizing quantization level and client selection in federal learning network

CN116471612BCN 116471612 BCN116471612 BCN 116471612BCN-116471612-B

Abstract

The invention provides a method for jointly optimizing quantization level and client selection in an energy collection type federal learning network. Firstly, constructing a federal learning network model carrying an energy acquisition module, secondly, equivalently replacing federal learning network training loss with expected distances of current and optimal weight models, thirdly, quantizing differential weights by a generalized quantization method and exploring the influence of the differential weights on the quantized optimal rate, then optimizing quantization level and client selection under the constraint of minimized energy causality to construct a non-convex mixed integer nonlinear optimization problem, and finally, constructing an MINLP problem as a convex optimization problem which can be solved based on a MINLP and other solvers.

Inventors

  • FAN LULU
  • NI ZHENGWEI
  • ZHANG CHAOYANG

Assignees

  • 浙江工商大学

Dates

Publication Date
20260512
Application Date
20230411

Claims (6)

  1. 1. A method for jointly optimizing quantization levels and client selection in an energy harvesting type federal learning network, comprising the steps of: Step 1, constructing a federal learning network model carrying an energy acquisition module; Step 2, equivalently replacing the federal learning network training loss problem with the expected distance problem of the current and optimal weight models; step 3, carrying out weight difference processing on the local weight vector, quantizing the difference weight by adopting a generalized quantization method, and exploring the influence of the weight difference on the quantized optimal rate; step 4, for the t+1st round of updating model, K clients are selected for training, and based on a quantization algorithm, the updated local weight vector is differentially quantized to obtain a new round of global weight vector; step 5, optimizing the quantization level and the client selection under the constraint of minimizing the causality of the energy, and constructing a non-convex mixed integer nonlinear optimization problem; Step 6, finally, converting the problem into a convex optimization problem, and solving the problem based on an MINLP solver, wherein in step 3, the problem is solved based on a loss function The L-shape of the product is smooth, -Strong convexity, unbiased and bounded client random gradient, can be obtained: ; Wherein, the , Represent the first The rate of the wheel-learning, E represents the local iteration number of the selected client, K represents the number of clients in each round, H represents the upper bound of random gradient descent of the client, Representing the upper bound of the client random gradient variance, Representing the degree of isomerism of the data set, Represent the first The weight vector of the +1 round, Representing a global optimal weight vector, T representing the number of rounds of global iteration during the training process, M being the number of selected clients in the T-th round, Represent the first A function of the number of time quantization bits; wherein the "loss function The L-shape of the product is smooth, Strong convexity refers in particular to a loss function Is L and the degree of bending is "Client side random gradient unbiased and bounded" means that the expectation of the client side k random gradient per round is equal to the true gradient and the random gradient variance is bounded; In step3, for the t-th round, we set the learning rate to be The boundaries with quantized federal learning satisfy: ; Wherein the method comprises the steps of E represents the selected local iteration number of the client, H represents the upper bound of random gradient descent of the client, Representing the upper bound of the client random gradient variance, Representing the degree of isomerism of the data set, Representing the loss after a T-round global training, Representing the optimal loss under global weight; if the weights are differentially quantized, federal learning can be overlaid on And as long as the loss function is satisfied The L-shape of the product is smooth, -Strong convexity, unbiased and bounded client random gradient, quantization levels for different rounds can be chosen arbitrarily; in step5, the reconstructed MINLP problem is: ; Wherein the method comprises the steps of I is an iteration variable, representing a pair From the slave To T terms for cumulative multiplication, L represents the rate of change of the loss function, Representing the degree of curvature of the loss function, E representing the number of local iterations of the selected client, H representing the upper bound of random gradient descent of the client, Representing the upper bound of the client random gradient variance, Representing the degree of isomerism of the data set, Represent the first The learning rate of the wheel is set to be equal to the learning rate of the wheel, The number of quantization bits representing the weight at round T; Due to And Is unchanged, and thus can convert the problem into a non-convex MINLP problem, namely: ; Wherein, the Indicating the degree of participation of the customer, Representing selection of clients at round t The training is carried out and the training is carried out, A function representing the number of quantization bits, Representing the length of the weight vector, The number of quantization bits representing the weight in the t-th round, Representing the number of bits quantized by client m in the t-th round, The time of the upload is indicated and, Representing the bandwidth allocated to each user, Representing a t-th round of clients Channel gain with the central parameter server, Representing the power spectral density of the noise in the network, Representing the energy consumed by the client in the t-th round of training, Representing the energy consumed by client m in a round of operation, Representing the energy collected by the t-th round, and in the step 1, constructing a federal learning network model: ; Wherein, the Representing a global loss function in federal learning, Representing the average loss of the mth client, and in step 2, equating the federal learning network training loss problem to the expected distance problem of the current and optimal weight models by: ; Wherein the method comprises the steps of Represented as a loss after T-round global training, Representing the optimal loss under the global weight, Representing the weight vector of the T-th round, L representing the rate of change of the loss function, Representing the desire between global weights and optimal weights after T rounds of training.
  2. 2. The method for selecting a client and a jointly optimized quantization level in an energy harvesting type federal learning network according to claim 1, wherein in step 3, a generalized quantization method is adopted to quantize differential weights, the method satisfies unbiasedness, and quantization errors satisfy two conditions by taking a function of the number of bits B as a boundary: ; Wherein the method comprises the steps of A method of quantization is indicated and, The function representing the number of quantization bits is a decreasing convex function.
  3. 3. The method for selecting a client and jointly optimizing quantization levels in an energy harvesting federal learning network according to claim 1, wherein in step 4, the method for updating weight vectors is as follows: The selection client is initialized and the client is selected, Indicating the degree of participation of the customer, Representing selection of clients at the t-th round Training, since only K clients are selected in each round, we have to have: ; in addition, to ensure the choice is unbiased, the number of times each client participates in training is set to be the same, namely: 。
  4. 4. A method for selecting a client and jointly optimizing quantization levels in an energy harvesting federal learning network according to claim 3, wherein in step 4, the central parameter server broadcasts to the clients of the round of selection, and the clients perform local updates.
  5. 5. The method for jointly optimizing quantization levels and client selection in an energy harvesting federal learning network according to claim 4, wherein in step 4, the local client has a new weight vector after E local iterations Using the same quantization level for all clients, the server aggregates the received local weight differences to generate a new weight vector: ; Wherein the method comprises the steps of Represent the first The weight vector of the wheel is used to determine, Represent the first The weight vector of the kth client in the round of training, Representing the set of clients for the t +1 round of selection, Representing the quantization method.
  6. 6. The method for selecting a joint optimization quantization level and a client in an energy harvesting type federal learning network according to claim 1, wherein in step 6, a non-convex MINLP problem is converted into a convex optimization problem, and a solution for solving the joint optimization quantization level and the client scheduling is found by solving the convex optimization problem: ; Wherein, the Indicating the degree of participation of the customer, Representing selection of clients at round t The training is carried out and the training is carried out, A function representing the number of quantization bits, Representing the length of the weight vector, The number of quantization bits representing the weight in the t-th round, The number of bits representing each weight that client m attempts to use in the t-th round, Representing the quantization level of client m in the case where round t is selected, Is the upper limit of the number of quantization bits, The time of the upload is indicated and, Representing the bandwidth allocated to each user, Representing a t-th round of clients Channel gain with the central parameter server, Representing the power spectral density of the noise in the network, Representing the energy consumed by the client in the t-th round of training, Representing the energy consumed by client m in a round of operation, Representing the energy collected at the t-th round.

Description

Method for jointly optimizing quantization level and client selection in federal learning network Technical Field The invention relates to the field of wireless communication, in particular to a design of a wireless federal learning system with an energy acquisition module. Background The federal learning network is a research hotspot in the current wireless communication field due to the characteristics of low privacy risk and low transmission cost and the combination of energy acquisition technology. The rapid increase of computing power of the processor and the explosive increase of mobile data mean that the client and the server need to have better information processing capability, and the limitation of wireless resources means that the resource allocation capability needs to be improved. Scheduling schemes are an important research direction in federal learning. In the training link. And bandwidth and energy resources are reasonably distributed among the clients, so that communication rounds are reduced, and communication quality is improved. Quantization techniques can greatly reduce the cost of wireless resources. Compared with the scheme without quantization, the same convergence speed can be maintained to a large extent as long as the design quantization is proper. If the model weights themselves are passed, the quantization levels should increase at a logarithmic rate. But the quantization level may remain unchanged if the weight difference is transferred instead. And introducing an energy acquisition module into the federal learning network to enable the client device to release the constraint of the wired power supply. The introduction of the energy harvesting module changes the resource allocation mechanism so that the parameters of each round in the system are related to the energy state and the energy arrival characteristics. The change of energy state has an important impact on the client scheduling and quantization level. In the present invention we allow the choice of the level of quantization for each round of federal learning independently from the current energy reserve and reveal the effect of each round of quantization on convergence speed. Disclosure of Invention The invention aims to solve the problem of how to provide high energy consumption for model uploading and local computing in a wireless network deploying federal learning. The invention configures an energy acquisition module for the client, uses the obtained energy to schedule the client to perform local calculation to obtain a weight vector, and adopts a generalized quantization method to quantize and upload the weight vector. In consideration of the energy limitation, a federal learning wireless network system with an energy acquisition module is designed, and a method for jointly optimizing the quantization level and the client selection under the system is provided. The technical scheme adopted by the invention is as follows: a method for jointly optimizing quantization level and client selection in an energy collection type federal learning network comprises the following steps: Step 1, constructing a federal learning network model carrying an energy acquisition module; Step 2, equivalently replacing the federal learning network training loss problem with the expected distance problem of the current and optimal weight models; Step 3, considering the influence of the weight vector on the quantization level, carrying out weight difference processing on the local weight vector, quantizing the difference weight by adopting a generalized quantization method, and exploring the influence of the weight difference on the quantization optimal rate; step 4, for the t+1st round of updating model, K clients are selected for training, and based on a quantization algorithm, the updated local weight vector is differentially quantized to obtain a new round of global weight vector; step 5, optimizing the quantization level and the client selection under the constraint of minimizing the causality of the energy, and constructing a non-convex mixed integer nonlinear optimization problem; And 6, finally, converting the problem into a convex optimization problem, and solving the problem based on an MINLP solver. Furthermore, on the basis of the technical scheme, each step can be further realized in the following specific mode. The method for constructing the federal learning network model carrying the energy acquisition module in the step 1 is as follows: Wherein, the Representing a global loss function in federal learning,Representing the number of clients that are to be included,Representing clientsIs a function of the loss of (2). Further, in the step 2, the method for equivalently replacing the federal learning network training loss problem with the expected distance problem of the current and optimal weight models is as follows: Wherein the method comprises the steps of Represented as a loss after T-round global training,Representing