CN113902021A - High-energy-efficiency clustering federal edge learning strategy generation method and device - Google Patents

High-energy-efficiency clustering federal edge learning strategy generation method and device Download PDF

Info

Publication number
CN113902021A
CN113902021A CN202111191599.8A CN202111191599A CN113902021A CN 113902021 A CN113902021 A CN 113902021A CN 202111191599 A CN202111191599 A CN 202111191599A CN 113902021 A CN113902021 A CN 113902021A
Authority
CN
China
Prior art keywords
edge
base station
learning
model
local
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111191599.8A
Other languages
Chinese (zh)
Other versions
CN113902021B (en
Inventor
秦晓琦
李艺璇
韩凯峰
马楠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing University of Posts and Telecommunications
China Academy of Information and Communications Technology CAICT
Original Assignee
Beijing University of Posts and Telecommunications
China Academy of Information and Communications Technology CAICT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing University of Posts and Telecommunications, China Academy of Information and Communications Technology CAICT filed Critical Beijing University of Posts and Telecommunications
Priority to CN202111191599.8A priority Critical patent/CN113902021B/en
Publication of CN113902021A publication Critical patent/CN113902021A/en
Application granted granted Critical
Publication of CN113902021B publication Critical patent/CN113902021B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

The invention discloses a clustering federal edge learning strategy generation method and a clustering federal edge learning strategy generation device with high energy efficiency, wherein the method comprises the following steps: s1, initializing an edge access strategy by the cloud center; s2, the edge base station solves the bandwidth resource allocation strategy of the access equipment and sends the initialization model to the access equipment; s3, calculating the precision of the received global model by the equipment, training the local model by adopting a layered migration strategy according to the global model and the local data, calculating the energy spent on uploading the local model, taking the difference value between the test precision and the energy consumption as the local profit, and uploading the local model and the local profit to the accessed edge base station; s4, the edge base station hierarchically aggregates the local model, calculates edge income by averaging local income of all access devices, and uploads the edge income to the cloud center; s5, the cloud center calculates the system profit according to the received feedback information of the edge base station, and adjusts an edge access strategy by adopting a deep reinforcement learning algorithm; and S6, repeating the above processes until convergence.

Description

High-energy-efficiency clustering federal edge learning strategy generation method and device
Technical Field
The invention relates to the technical field of data processing, in particular to a clustering federal edge learning strategy generation method and device with high energy efficiency.
Background
Data security has become a key issue for the continuous development of artificial intelligence technology. Traditional machine learning techniques are centralized and collect device data to a processing center for centralized training, however, this may lead to leakage of user data privacy.
Federal learning is a promising distributed machine learning architecture, and with the improvement of computing capacity of equipment, the equipment can train a local model by locally using collected data, and then only needs to upload the local model to a processing center for model aggregation, so that the direct uploading of original data is avoided, and the data privacy is greatly protected.
In real life, data between devices may present non-independent and identically distributed characteristics, which presents a challenge for federal learning to train a unified global model. It is therefore of great interest to study how federal learning adapts to the data on each device. Currently, some research has been proposed to personalize federal learning.
The personalized federal learning comprises federal transfer learning and federal meta learning, and the essence of the personalized federal learning is that a basic global model shared by all devices is obtained firstly, and then the global model is finely adjusted on each device according to local data to adapt to personalized data characteristics. Different personalized federal learning strategies each have drawbacks. Due to the facts that the federal transfer learning and the federal meta learning need to obtain a global model covering most of characteristics firstly and then carry out personalization, the federal transfer learning and the federal meta learning are only suitable for data with weak isomerism and cannot process the personalization problem of a system with strong isomerism data.
Multitask federated learning is also an effective method for solving the personalized federated learning, and the similarity of different equipment models is quantified by calculating a correlation matrix, and then heterogeneous data is used as different learning targets, so that the multitask learning is performed. The federal multi-task learning is only suitable for convex problems or double convex problems, is difficult to be expanded to non-convex problems such as common neural networks and has limitation. In addition, most of these personalization methods are suitable for outputting data with different labels, for example, each device only has a subset of all labels, and cannot be suitable for data with different condition distributions and obvious cluster structures.
The clustering federation learning can effectively solve the problems, and can capture the clustering structure among data, so that a plurality of models are aggregated according to data distribution to meet the heterogeneous data characteristics among equipment, and the learning accuracy is greatly improved. Due to privacy of federal learning, data distribution on devices is unknown, which presents a significant challenge to clustering. Theoretical analysis can show that when the distance between the learning models is smaller, the data distribution of the learning models is closer, so that under the condition of not uploading original data, the clustering federation learning mostly adopts the model distance to measure the data similarity on different devices. Common indexes for measuring the model distance are Euclidean distance, cosine distance and the like. However, some techniques may infer data information at the device from the local model, thereby causing data privacy to be compromised. The model nonlinear encryption method can solve the problem well, but the distance between models after nonlinear encryption is not in proportion to the distance of an original model, so that the local model distance clustering is used, although the calculation complexity is low, in this case, the similarity between data cannot be judged through the models after encryption, and the clustering method is invalid, so that the method is not a widely applicable method. In addition, most of the existing clustering federation learning only considers the statistical heterogeneity of data, and neglects the resource limitation and communication bottleneck problems of the system. Meanwhile, the research only considers the scene of a single base station and lacks the expansion of multiple base stations. For energy-limited devices, communication overhead is not negligible, spectrum resources provided by a single base station are limited, and for devices with poor channel conditions, uploading a local model consumes a large amount of device energy, thereby reducing learning performance in a training cost budget.
Traditional federal learning requires devices to upload local models to the cloud for aggregation through a wide area network, the battery capacity of the devices is often limited, and multiple communication iterations of federal learning and huge communication overhead in each iteration consume a large amount of transmission energy, thereby reducing learning performance at a given energy budget. Multi-access edge computing (MEC) technology is a promising distributed computing framework that can support the needs of many low-latency, low-energy applications, MECs offloading delay-sensitive and compute-intensive tasks to the edge, enabling real-time and energy-efficient. The Federal edge learning utilizes the advantages of MEC, a plurality of base stations are added between the cloud and the equipment to further assist training, and the equipment uploads the local model to the edge base station for aggregation. The communication overhead of the equipment and the cloud transmitted through the wide area network is greatly reduced, and in addition, the system is enabled to realize high energy efficiency and high precision under the condition of non-independent and same distribution of data through the overall coordination of the edge base station and the equipment. In the multi-base-station federated learning architecture, mostly only training cost such as time and energy consumption is considered, opportunities and challenges brought to a multi-base-station scene by statistical heterogeneity are not considered, and joint optimization research aiming at the training cost and the learning performance is lacked.
Disclosure of Invention
Aiming at the defects of the prior art, under the scene of a multi-edge base station, the invention jointly considers the data distribution and the energy consumption cost of equipment, finds a cross point for counting heterogeneous and communication bottlenecks, designs an edge access strategy and a resource allocation strategy with high energy efficiency and high precision from the perspective of system benefits, and provides a clustering federal edge learning strategy generation method and a clustering federal edge learning strategy generation device with high energy efficiency.
In order to achieve the above purpose, the invention provides the following technical scheme:
in a first aspect, the invention provides a high-energy-efficiency clustering federal edge learning strategy generation method, which comprises the following steps:
s1, initializing an edge access strategy by the cloud center;
s2, the edge base station uses the convex optimization method to solve the bandwidth resource allocation strategy of the access device, and sends the initialization model to the access device;
s3, the equipment calculates the accuracy of the received global model on a local test data set, trains the local model by adopting a layered federal migration method according to the global model and the local training data, calculates the energy consumed by uploading the local model, takes the difference value between the test accuracy and the energy consumption as local income, and uploads the local model and the local income to the accessed edge base station;
s4, the edge base station hierarchically aggregates the local models, calculates edge income by averaging local income of all access devices, and uploads the edge income to the cloud center;
s5, the cloud center calculates the system profit according to the received feedback information of the edge base station, and adjusts an edge access strategy by adopting a deep reinforcement learning algorithm;
and S6, repeating the above processes until convergence.
Further, in step S1, the access policy a between the device and the edge base stationijIs a binary variable, i.e. if device i communicates with edge base station j, then aij1, otherwise, aijEach device accesses one edge base station, 0.
Further, the convex optimization method in step S2 specifically includes: for edge base station j and access equipment cluster thereof
Figure BDA0003301339640000031
Optimal bandwidth allocation beta for resource allocation sub-problem given edge access policyijThe calculation formula is as follows:
Figure BDA0003301339640000041
wherein h isijRepresenting the channel gain, p, between device i and edge server jiModel upload Power, N, representing device i0Is highPower spectral density of the noise, betaijBjFor the bandwidth resources divided by the device i accessing the edge base station j, the shared bandwidth of the device accessing the edge base station j is BjIs used for communication over a common frequency spectrum,
Figure BDA0003301339640000042
aijindicating the access policy of the device to the edge base station, betaijRepresenting the proportion of bandwidth allocated to device i.
Further, the device is based on the received global model θjUsing local data
Figure BDA0003301339640000043
Training is carried out, and the loss function formula of the device i is as follows:
Figure BDA0003301339640000044
the device updates the local model omega using a gradient descent methodiThe formula is as follows:
Figure BDA0003301339640000045
wherein eta is a learning step length, and eta is more than or equal to 0;
step S3, training a local model by adopting a layered federal transfer learning strategy, dividing a neural network into a basic characteristic layer and an individual characteristic layer, wherein the layered federal transfer learning strategy comprises the following specific processes:
s301, calculating the average learning precision of each edge base station after a certain turn according to the following formula:
Figure BDA0003301339640000046
s302, device basic feature layer model with higher average precision
Figure BDA0003301339640000047
And a personality trait layer model
Figure BDA0003301339640000048
Uploading to an accessed edge base station, uploading a basic characteristic layer model to equipment with lower average precision, and locally updating the individual characteristic layer model in the equipment, wherein the formula is as follows:
Figure BDA0003301339640000049
wherein,
Figure BDA00033013396400000410
for the local personality trait layer model of device i,
Figure BDA00033013396400000411
is migrating a device set.
And S303, the edge base station aggregates the basic feature layer models of all the devices and aggregates the individual feature layer models of the non-migration devices, the edge base station issues the aggregated basic feature layer models to all the access devices and issues the individual feature layer models to the non-migration devices, and the devices perform the updating according to the received models and iterate until convergence.
Further, in step S3, the learning accuracy g of the global model on the local test data setijAs an index for measuring the performance of the global model on the edge base station j, the learning performance gain G of the system is the average accuracy of all the devices, and the formula shows:
Figure BDA0003301339640000051
further, the energy E consumed by the device i uploading the local model in step S3ijThe formula is as follows:
Figure BDA0003301339640000052
Tijthe transmission delay for device i to upload the local model to the edge base station is given by the following formula:
Figure BDA0003301339640000053
s represents the size of the local model, rijFor the transmission rate of the upload model of device i, the formula is as follows:
Figure BDA0003301339640000054
hijrepresenting the channel gain, p, between device i and edge server jiModel upload Power, N, representing device i0Power spectral density, beta, representing gaussian noiseijBjFor the bandwidth resources divided by the device i accessing the edge base station j, the shared bandwidth of the device accessing the edge base station j is BjIs used for communication over a common frequency spectrum,
Figure BDA0003301339640000055
aijindicating the access policy of the device to the edge base station, betaijRepresenting the proportion of bandwidth allocated to device i. Further, in step S4, before hierarchical aggregation, the edge base station aggregates all received local models, and the formula is as follows:
Figure BDA0003301339640000056
wherein,
Figure BDA0003301339640000057
for all clusters of devices accessing the edge base station j, omegaiIs a local model.
After a certain round of training, executing a layered federal migration learning strategy, and hierarchically aggregating received local models by an edge base station, wherein the specific method comprises the following steps: the edge base station aggregates basic characteristic layer models of all the devices to ensure the generalization performance of the models, and aggregates individual characteristic layer models of the non-migration devices to eliminate the influence of non-independent and same distributed data among the devices, and the formula is as follows:
Figure BDA0003301339640000058
Figure BDA0003301339640000061
wherein,
Figure BDA0003301339640000062
the base feature layer global model for the device accessing edge base station j, shared for all devices,
Figure BDA0003301339640000063
for the personality-level global model, for non-migration device sharing,
Figure BDA0003301339640000064
is a non-migrating device cluster accessing the edge base station j.
Further, the formula of the system revenue function in step S5 is as follows:
Figure BDA0003301339640000065
where μ is a continuous variable and μ ∈ [0,1 ]]For adjusting the trade-off between learning performance and transmission power consumption, GmaxAnd EmaxThe highest accuracy and the maximum energy consumption of the system can be achieved.
Further, in step S5, the edge access policy is adjusted by deep reinforcement learning, and the specific process of deep reinforcement learning is as follows:
s501, describing the edge correlation problem as a Markov process
Figure BDA0003301339640000066
The specific details are as follows:
(1) status of state
Figure BDA0003301339640000067
In the k-th round, the state is defined as S (k) { S ═ S1(k),S2(k),…,SN(k) Each item Si(k) Is defined as:
Si(k)＝{Ai(k-1),βij(k),Δi(k)}
wherein, Deltai(k) Indicates whether the learning accuracy is improved, i.e., Δ, compared to the k-1 roundi(k) 1 stands for improved accuracy, whereas Δi(k)＝0；
(2) Movement of
Figure BDA0003301339640000068
In the k-th round, the actions associate policies for the edge of each device:
A(k)＝{A1(k)，A2(k)，…，AN(k)}
wherein each item Ai(k) Can be expressed as:
Ai(k)＝{aij(k)}
(3) reward
Figure BDA0003301339640000069
Set the reward as the objective function:
Figure BDA00033013396400000610
s502, selecting DQN as a basic frame, optimizing an algorithm by combining dulling DQN and double DQN, solving an edge access problem by using D3QN, approximating a Q value function Q (S, A; theta) by a neural network with a parameter theta to represent a mapping relation between environment and action, and obtaining the output of the neural network through a Bellman equation:
Figure BDA00033013396400000611
wherein, S ', A ', theta ' are the state, action and corresponding parameters of the next time slot respectively;
two Q networks of the same structure but different parameters are used in DQN to improve the stability of the algorithm, one is the current Q network with the latest parameters to evaluate the current state-action cost function, the other is the target Q network with the past round parameters and keep the Q value unchanged for a period of time, the Q value of the current Q network is used as the input of the neural network, the goal of DQN is to minimize the difference between the two Q networks and define it as the loss function of DQN:
L(θ)＝E[(y-Q(S，A；θ))2]
s503, selecting the action corresponding to the maximum Q value in the current Q network by adopting a DDQN algorithm:
Figure BDA0003301339640000071
and then bringing the selected action into the target Q network to calculate a Q value:
y＝R(S，A)+γQ'(φ(S'),Amax(S'；θ)；θ')
s504, using blanking DQN to optimize the network structure, and dividing the network into two parts, namely a value function V (S, theta, alpha) only related to the state and a potential function A (S, A, theta, beta) related to both the state and the action, wherein theta is a common parameter of the two networks, alpha is a parameter unique to the value function, beta is a parameter unique to the potential function, and Q is the sum of the two functions:
Q(S,A,θ,α,β)＝V(S,θ,α)+A(S,A,θ,β)。
in a second aspect, the present invention provides an energy-efficient clustered federal edge learning policy generation apparatus, including a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer processor implements the above-mentioned energy-efficient clustered federal edge learning policy generation method when executing the computer program.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention jointly considers the practical bottleneck of carrying out the federal learning in reality, namely the non-independent and distributed characteristics of the equipment data and the communication and energy limits of the equipment, and most researches only consider a single problem.
2. The invention considers the communication overhead of the equipment, and widens the federal learning of the traditional single base station to a multi-base station scene. Different from the situation that only the communication bottleneck problem is solved in a multi-base station scene, the invention jointly considers the heterogeneity and the channel state of data distribution from the perspective of system income, and designs the edge access strategy and the resource allocation strategy with high precision and high energy efficiency.
3. In order to increase the universality of the algorithm, the invention considers the data privacy problem of federal learning, particularly some technologies can deduce the data at the equipment end from the model uploaded by the equipment, and the nonlinear privacy encryption is an algorithm for further protecting the data privacy, so that the common model distance clustering method is invalid. The invention designs deep reinforcement learning, adaptively explores an edge access strategy according to edge feedback information, and protects data privacy. Meanwhile, in order to increase the expansibility of the algorithm and reduce the complexity of the algorithm, the resource allocation problem is decoupled to the edge base station to be solved independently.
4. The invention considers the condition that the equipment with inconsistent data distribution is accessed to the same edge base station, and designs the layered transfer learning to further improve the learning performance. Analysis can be carried out, and the layered migration strategy designed by the invention does not consume extra energy.
Drawings
In order to more clearly illustrate the embodiments of the present application or technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings can be obtained by those skilled in the art according to the drawings.
Fig. 1 is a clustered federal edge learning system architecture provided in an embodiment of the present invention.
Detailed Description
For a better understanding of the present solution, the method of the present invention is described in detail below with reference to the accompanying drawings.
The invention provides a clustering federal edge learning strategy generation method with high energy efficiency, which comprises the following steps:
s1, initializing an edge access strategy by the cloud center;
s2, the edge base station uses the convex optimization method to solve the bandwidth resource allocation strategy of the access device, and sends the initialization model to the access device;
s3, the equipment calculates the precision of the received global model on a local test data set, trains the local model by adopting a layered federal migration method according to the global model and the local training data, calculates the energy consumed by uploading the local model, takes the difference value between the test precision and the energy consumption as local income, and uploads the local model and the local income to the accessed edge base station;
s4, the edge base station hierarchically aggregates the local models, calculates edge income by averaging local income of all access devices, and uploads the edge income to the cloud center;
s5, the cloud center calculates the system profit according to the received feedback information of the edge base station, and adjusts an edge access strategy by adopting a deep reinforcement learning algorithm;
and S6, repeating the above processes until convergence.
The invention considers a clustering federal edge learning framework under a multi-base-station scene, and as shown in fig. 1, the clustering federal edge learning framework consists of a cloud center S, M edge base stations and N devices. In a network, to
Figure BDA0003301339640000091
As a set of edge base stations, the edge base stations,
Figure BDA0003301339640000092
for the number of edge base stations,
Figure BDA0003301339640000093
Figure BDA0003301339640000094
is a set of devices that are to be considered,
Figure BDA0003301339640000095
is the number of devices. For each device
Figure BDA0003301339640000096
Collecting and storing training data sets
Figure BDA0003301339640000097
Wherein xinStore the sample for the nth of device i, yinIs xinThe corresponding label of (a) to (b),
Figure BDA0003301339640000098
is the amount of training data for device i. Training data of different devices are acquired from different data sources, so that the training data of federal learning are not independently and uniformly distributed.
In clustered federated edge learning, the goal of the system is to learn multiple models to satisfy heterogeneous data on a device. The federal learning training procedure includes the following steps:
the edge base station sends an initial global model to the equipment;
the device receives the global model thetajUsing local data
Figure BDA0003301339640000099
And (5) training. The loss function for device i is defined as:
Figure BDA00033013396400000910
the device updates the local model omega using a gradient descent methodiAs follows:
Figure BDA00033013396400000911
wherein eta is the learning step length, and eta is more than or equal to 0.
Uploading the updated local model to an accessed edge base station through a wireless link;
the edge base station aggregates all received local models as follows:
Figure BDA00033013396400000912
wherein,
Figure BDA00033013396400000913
is a cluster of devices that all access edge base station j.
The above process is repeated until the model converges.
Access strategy a of equipment and edge base stationijIs a binary variable, i.e. if device i communicates with edge base station j, then aij1, otherwise, aij0. Each device can access only one edge base station, so the invention has:
Figure BDA0003301339640000101
the invention leads the learning precision g of the global model on the local data setijAs an index for measuring the global model performance at the edge base station j, the learning performance gain G of the system can be regarded as the average accuracy of all devices, as follows:
Figure BDA0003301339640000102
it is worth noting that the access of the devices with non-independent and equally distributed training data to the same edge base station has a negative impact on the learning performance, so statistical heterogeneity is a key issue that is not negligible when designing the edge access policy.
For the uploading process of the local model, the invention adopts orthogonal frequency division multiple access (orthogona)l frequency division multiple access, OFDMA), which is also easily extendable to other communication systems. All devices accessing the edge base station j share the available bandwidth BjCommunicate over a common frequency spectrum, betaijRepresenting the proportion of bandwidth allocated to device i. Then the invention has:
Figure BDA0003301339640000103
Figure BDA0003301339640000104
through the analysis, the bandwidth resource divided by the equipment i accessing the edge base station j is betaijBj. The transmission rate of the device i upload model may be expressed as follows:
Figure BDA0003301339640000105
wherein h isijRepresenting the channel gain, p, between device i and edge server jiModel upload Power, N, representing device i0Representing the power spectral density of gaussian noise. Let S denote the size of the local model, the transmission delay of the device i uploading the local model to the edge base station may be expressed as follows:
Figure BDA0003301339640000106
the energy consumed by device i to upload the local model may then be expressed as follows:
Figure BDA0003301339640000107
the invention takes the average transmission energy consumption of all the devices as the communication cost of the Federal learning system, obviously, the communication cost can be easily expanded to other resources, such as training time delay and the like. The communication cost of the system can be expressed as follows:
Figure BDA0003301339640000108
from the above analysis, both the edge access policy and the bandwidth resource allocation policy affect the device energy consumption. Therefore, the communication cost should also be considered when designing the edge access policy.
In order to improve the learning precision while saving the communication cost, the invention quantifies the integral performance of the federal learning by the system benefit. The present invention defines the system benefits as follows:
Figure BDA0003301339640000111
where μ is a continuous variable and μ ∈ [0,1 ]]And the method is used for adjusting the balance relation between the learning performance and the transmission energy consumption. GmaxAnd EmaxThe highest accuracy and the maximum energy consumption of the system can be achieved. The purpose of the regularization is to mitigate the impact of the two different orders of magnitude on the strategy.
The aim of the invention is to find an edge access strategy and a resource allocation strategy to maximize the system benefit. The optimization problem can be expressed as follows:
max P
s.t.
Figure BDA0003301339640000112
Figure BDA0003301339640000113
Figure BDA0003301339640000114
Figure BDA0003301339640000115
in the objective function, aijBeing binary variables, betaijIs a continuous variable. This optimization problem can be expressed as a mixed integer nonlinear programming problem (MINLP).
Due to privacy of federal learning, statistical distribution of device data is not available, so it is very difficult to directly obtain a global optimal solution. Meanwhile, in order to prevent original data information from being obtained from local model parameters uploaded by equipment, federal learning is often combined with a nonlinear privacy encryption method. In view of the problem and in order to increase the universality of the proposed algorithm, the invention uses deep reinforcement learning to adaptively explore an edge access strategy in a multi-base station scene according to edge feedback information, and can maximize the profit of the system in a way of protecting data privacy without data exchange.
The deep reinforcement learning can convert different types of variables into the same type for unified solution in modes of discretization continuous variables or continuous discrete variables and the like. However, as the solution variables increase, deep reinforcement learning easily falls into a locally optimal solution, resulting in unsatisfactory results. Therefore, the invention decouples the original problem into two subproblems to solve, which respectively are as follows: the edge association problem is associated with the resource allocation problem given the edge access policy. For the edge association subproblem, deep reinforcement learning is deployed at the cloud end to adaptively adjust the access strategy between the edge base station and the equipment. The resource allocation sub-problem is related to the edge access problem, so that the resource allocation strategy is decoupled to each edge base station to be solved independently under the condition of giving the edge access strategy, the complexity of the algorithm is reduced, and the expansibility of the algorithm is increased.
The invention observes that when the edge access strategy is fixed, the learning performance of the system is determined accordingly, so that the optimization problem can be simplified into the problem of how to allocate communication resources to minimize the energy consumption of uploading. And the bandwidth resource of each base station is independently determined by the base station and is independent of other edge base stations. Therefore, the resource allocation problem of the multi-edge base station can be decomposed into M sub-problems, and the sub-problems are solved separately on each edge base station. For each edge base station, the following problem needs to be solved:
Figure BDA0003301339640000121
Figure BDA0003301339640000122
Figure BDA0003301339640000123
wherein,
Figure BDA0003301339640000124
for accessing clusters of devices at edge base station j, NjAs a group of devices
Figure BDA0003301339640000125
The number of devices in (1).
It is clear that the above problem is a convex problem. Because of the variable betaijConvex in the feasible region and affine in all constraints.
The invention uses the commonly used Karush-Kuhn-Tucker (KKT) condition to obtain the analytic solution of bandwidth allocation, and has the following theorem.
Theorem 1: for edge base station j and its training equipment cluster
Figure BDA0003301339640000126
Optimal bandwidth allocation beta for resource allocation sub-problem given edge access policyijCan be expressed as follows:
Figure BDA0003301339640000127
the proof process of theorem 1 is as follows:
the convex problem can be solved by a lagrange multiplier method, and the lagrange equation of the sub-problem objective function can be expressed as follows:
Figure BDA0003301339640000131
where λ is the lagrange multiplier of the convex problem constraint. To solve the lagrangian equation, the invention calculates its KKT condition:
Figure BDA0003301339640000132
Figure BDA0003301339640000133
by solving the above equation, one can obtain:
Figure BDA0003301339640000134
based on this, the present invention can obtain the bandwidth allocation and the expression of the lagrangian multiplier, and then the present invention has:
Figure BDA0003301339640000135
Figure BDA0003301339640000136
meanwhile, according to the KKT condition, the invention comprises the following components:
Figure BDA0003301339640000137
thus, it is possible to obtain:
Figure BDA0003301339640000138
by the above formula, the present invention can solve the bandwidth resource allocation variable, which can be expressed as:
Figure BDA0003301339640000141
by theorem 1, the invention can effectively solve the problem of communication resource allocation, and for a given edge access strategy, the invention has the optimal bandwidth resource allocation strategy under the condition, and forms a one-to-one corresponding relation, thereby reducing the difficulty in solving the original problem.
For the edge access problem, the traditional method needs to obtain all information and then solve, but due to the privacy of federal learning, the traditional method is impossible. Deep reinforcement learning is an algorithm that does not require any a priori information by constantly exploring the environment. The invention designs a deep reinforcement learning method capable of adaptively adjusting an edge access strategy according to feedback information of an edge base station. The edge correlation problem may be described as a Markov process
Figure BDA0003301339640000142
The specific details are as follows:
(1) status of state
Figure BDA0003301339640000143
In the k-th round, the cloud can only observe feedback information from the edge base station to the edge access strategy in the previous round, so the present invention defines the state as S (k) { S ═ S1(k),S2(k),…,SN(k) Each item Si(k) Can be defined as:
Si(k)＝{Ai(k-1),βij(k),Δi(k)}
wherein, Deltai(k) Indicates whether the learning accuracy is improved, i.e., Δ, compared to the k-1 roundi(k) 1 stands for improved accuracy, whereas Δi(k)＝0。
(2) Movement of
Figure BDA0003301339640000144
In the k-th round, the actions associate policies for the edge of each device:
A(k)＝{A1(k),A2(k),…,AN(k)}
wherein each item Ai(k) Can be expressed as:
Ai(k)＝{aij(k)}
(3) reward
Figure BDA0003301339640000145
The reward is the guideline of the strategy, so the invention sets the reward as the objective function:
Figure BDA0003301339640000146
since the edge base station is not aware of all possible subsequent states and optimal actions, the present invention uses a model-free deep reinforcement learning paradigm to update the edge access policy. Meanwhile, in order to handle large state space and discrete type actions, the invention selects a Deep Q Network (DQN) as a basic framework, and optimizes the algorithm of the invention in combination with dulling DQN and double DQN, using D3QN to solve the edge access problem.
DQN is a value-based algorithm, the Q value function Q (S, A; theta) is approximated by a neural network with parameters theta, representing the mapping relation between environment and action, the output of the neural network can be obtained by Bellman equation, the invention has:
Figure BDA0003301339640000151
wherein, S ', a ', θ ' are the state, action and corresponding parameters of the next time slot, respectively.
Two Q networks with the same structure but different parameters are used in the DQN to improve the stability of the algorithm. One is the current Q network with the latest parameters to evaluate the current state-action cost function. The other is a target Q network with past round parameters and keeping the Q constant for a period of time. The invention takes the Q value of the current Q network as the input of the neural network. Obviously, the goal of DQN is to minimize the difference between the two Q networks and define it as a loss function of DQN. The invention comprises the following steps:
L(θ)＝E[(y-Q(S,A；θ))2]
in order to meet the non-independent and same distribution characteristics of data in the Markov process, the DQN adopts an empirical playback strategy to reduce the time correlation among samples and ensure the stability of the algorithm. However, the target values of DQN are all obtained directly by the greedy method, which results in over-estimation and large bias. To address this problem, the present invention introduces a DDQN algorithm that avoids over-estimation by decoupling the selection of target actions and the evaluation of the current state. Different from the action of selecting the maximum Q value in the target Q network in the DQN, the action corresponding to the maximum Q value in the current Q network is selected by the DDQN, and the method comprises the following steps:
Figure BDA0003301339640000152
and substituting the selected action into the target Q network to calculate the Q value, the invention comprises the following steps:
y＝R(S，A)+γQ'(φ(S'),Amax(S'；θ)；θ')
meanwhile, in order to converge more quickly, the invention uses dulling DQN to optimize the network structure and divides the network into two parts, namely a value function V (S, theta, alpha) only related to the state and a potential function A (S, A, theta, beta) related to both the state and the action, wherein theta is a common parameter of the two networks, alpha is a unique parameter of the value function, and beta is a unique parameter of the potential function. The Q value can be regarded as the sum of these two functions, and the invention has:
Q(S，A，θ,α，β)＝V(S，θ,α)+A(S，A，θ，β)
the dulling DQN can better evaluate the policy, thereby speeding up the convergence of the network.
It is worth noting that the edge access policy obtained by the cloud center directly changes the access relationship between the edge base station and the device, and further guides the communication resource allocation policy on the edge base station, thereby affecting the system learning performance and the device energy consumption.
Considering that the system may have energy consumption balance, the devices with different data distribution access the same edge base station, the invention utilizes the advantage of transfer learning and designs a layered federal transfer learning strategy. The invention can divide the neural network into a basic characteristic layer and an individual characteristic layer. The basic characteristic layer has common characteristics of most data, and the individual characteristic layer captures unique properties of different data. The layered federal migration learning of the present invention is described in detail as follows:
(1) identifying a migration device: the invention calculates the average learning precision of each edge base station after a certain turn:
Figure BDA0003301339640000161
in the invention, the equipment with lower average precision is regarded as the equipment with further improved precision, obviously, the equipment precision different from most data distribution in the equipment cluster of the edge base station is lower than the average precision. For convenience, the present invention is referred to collectively as migration devices, and the other devices are referred to as non-migration devices.
(2) Layered federal migration learning: non-migration equipment models the basic characteristic layer thereof
Figure BDA0003301339640000162
And a personality trait layer model
Figure BDA0003301339640000163
And uploading to the accessed edge base station. The migration equipment only uploads the basic characteristic layer model, and the individual characteristic layer model is updated locally in the equipment, and the invention comprises the following steps:
Figure BDA0003301339640000164
the edge base station aggregates the basic characteristic layer models of all the devices to ensure the generalization performance of the models, and aggregates the individual characteristic layer models of the non-migration devices to eliminate the influence of the non-independent co-distributed devices. Then the invention has:
Figure BDA0003301339640000165
Figure BDA0003301339640000166
(3) and the edge base station issues the aggregated basic layer model to all the access devices, and issues the individual layer model to the non-migration device. The device performs the above updating again according to the received model, and iterates until convergence.
The layered migration learning strategy provided by the invention does not consume extra energy, because the same as the traditional federal learning, the equipment updates each layer of model when training locally, and the difference is that the non-migration equipment needs to upload each layer of model when uploading, and the migration equipment only needs to upload the individual characteristic layer model to the accessed edge base station, which reduces the size of the uploaded model. But the base layer model accounts for the majority of all layers, so the invention ignores this reduced energy consumption when calculating the energy consumption.
In summary, the invention provides a clustering federal edge learning strategy generation method with high energy efficiency, which comprises the following steps:
firstly, in order to achieve the purpose of high-efficiency learning of a federal system, the learning performance is used as the system harvest, and the communication energy consumption is used as the system cost, so that a system profit function is obtained. In order to research the system profit optimization problem in the clustering federated edge learning network, the invention jointly considers the heterogeneous characteristics of communication conditions and data, realizes high energy efficiency while ensuring the learning performance, and quantifies the problem into a mixed integer nonlinear programming (MINLP) problem.
Secondly, in order to effectively solve the problem of maximizing the system yield, the invention observes that after the edge access strategy is determined, the original problem can be regarded as a resource allocation problem aiming at high energy efficiency, so the original problem is decomposed into two sub-problems, namely the edge access problem and the resource allocation problem of the given edge access strategy, and an effective iterative optimization algorithm is designed according to the two sub-problems. For the edge access sub-problem, in order to enhance the privacy of the federal learning data and be better suitable for the model nonlinear encryption algorithm, the invention explores the edge access strategy by deep reinforcement learning. In the sub-problem of resource allocation, in order to reduce the complexity of the algorithm, a convex optimization algorithm is adopted to solve the resource allocation strategy.
Finally, due to balance of energy consumption, devices with different data distribution may be accessed to the same base station for training together, and in consideration of the situation, the invention provides a layered federal transfer learning strategy, so that the learning precision is further improved under the condition of not additionally consuming energy.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: it is to be understood that modifications may be made to the technical solutions described in the foregoing embodiments, or equivalents may be substituted for some of the technical features thereof, but such modifications or substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A clustering federal edge learning strategy generation method with high energy efficiency is characterized by comprising the following steps:
s1, initializing an edge access strategy by the cloud center;
s2, the edge base station uses the convex optimization method to solve the bandwidth resource allocation strategy of the access device, and sends the initialization model to the access device;
s3, the equipment calculates the accuracy of the received global model on a local test data set, trains the local model by adopting a layered federal migration method according to the global model and the local training data, calculates the energy consumed by uploading the local model, takes the difference value between the test accuracy and the energy consumption as local income, and uploads the local model and the local income to the accessed edge base station;
s4, the edge base station hierarchically aggregates the local models, calculates edge income by averaging local income of all access devices, and uploads the edge income to the cloud center;
s5, the cloud center calculates the system profit according to the received feedback information of the edge base station, and adjusts an edge access strategy by adopting a deep reinforcement learning algorithm;
and S6, repeating the above processes until convergence.
2. The method for generating the energy-efficient clustering federated edge learning strategy of claim 1, wherein in step S1, the access strategy a between the device and the edge base stationijIs a binary variable, i.e. if device i communicates with edge base station j, then aij1, otherwise, aijEach device accesses one edge base station, 0.
3. The energy-efficient clustering federal edge learning strategy generation method according to claim 1, wherein the convex optimization method in step S2 is specifically: for edge base station j and access equipment cluster thereof
Figure FDA0003301339630000011
Optimal bandwidth allocation beta for resource allocation sub-problem given edge access policyijThe calculation formula is as follows:
Figure FDA0003301339630000012
wherein h isijRepresenting the channel gain, p, between device i and edge server jiModel upload Power, N, representing device i0Power spectral density, beta, representing gaussian noiseijBjFor the bandwidth resources divided by the device i accessing the edge base station j, the shared bandwidth of the device accessing the edge base station j is BjIs used for communication over a common frequency spectrum,
Figure FDA0003301339630000021
aijindicating the access policy of the device to the edge base station, betaijRepresenting the proportion of bandwidth allocated to device i.
4. The energy-efficient clustering federated edge learning strategy generation method of claim 1, wherein in step S3, the device follows the received global model θjUsing local data
Figure FDA0003301339630000022
Training is carried out, and the loss function formula of the device i is as follows:
Figure FDA0003301339630000023
the device updates the local model omega using a gradient descent methodiThe formula is as follows:
Figure FDA0003301339630000024
wherein eta is a learning step length, and eta is more than or equal to 0;
step S3, after training for a certain turn, training a local model by adopting a layered federal migration learning strategy, dividing a neural network into a basic characteristic layer and an individual characteristic layer, wherein the specific process of the layered federal migration learning strategy is as follows:
s301, calculating the average learning precision of each edge base station after a certain turn according to the following formula:
Figure FDA0003301339630000025
s302, device basic feature layer model with higher average precision
Figure FDA0003301339630000026
And a personality trait layer model
Figure FDA0003301339630000027
Uploading to an accessed edge base station, uploading a basic characteristic layer model to equipment with lower average precision, and locally updating the individual characteristic layer model in the equipment, wherein the formula is as follows:
Figure FDA0003301339630000028
wherein,
Figure FDA0003301339630000029
for the local personality trait layer model of device i,
Figure FDA00033013396300000210
is migrating a device set.
And S303, the edge base station aggregates the basic feature layer models of all the devices and aggregates the individual feature layer models of the non-migration devices, the edge base station issues the aggregated basic feature layer models to all the access devices and issues the individual feature layer models to the non-migration devices, and the devices perform the updating according to the received models and iterate until convergence.
5. The method for generating an energy-efficient clustered federated edge learning strategy according to claim 1, wherein in step S3, the learning accuracy g of the global model on the local test data setijAs an index for measuring the performance of the global model on the edge base station j, the learning performance gain G of the system is the average accuracy of all the devices, and the formula shows:
Figure FDA0003301339630000031
6. the chair of claim 1The method for generating the energy efficiency clustering federal edge learning strategy is characterized in that in the step S3, the equipment i uploads the energy E consumed by the local modelijThe formula is as follows:
Figure FDA0003301339630000032
Tijthe transmission delay for device i to upload the local model to the edge base station is given by the following formula:
Figure FDA0003301339630000033
s represents the size of the local model, rijFor the transmission rate of the upload model of device i, the formula is as follows:
Figure FDA0003301339630000034
hijrepresenting the channel gain, p, between device i and edge server jiModel upload Power, N, representing device i0Power spectral density, beta, representing gaussian noiseijBjFor the bandwidth resources divided by the device i accessing the edge base station j, the shared bandwidth of the device accessing the edge base station j is BjIs used for communication over a common frequency spectrum,
Figure FDA0003301339630000035
aijindicating the access policy of the device to the edge base station, betaijRepresenting the proportion of bandwidth allocated to device i.
7. The method for generating an energy-efficient clustering federated edge learning strategy according to claim 1, wherein in step S4, before hierarchical aggregation, the edge base station aggregates all received local models, and the formula is as follows:
Figure FDA0003301339630000036
wherein,
Figure FDA0003301339630000037
for all clusters of devices accessing the edge base station j, omegaiIs a local model.
After a certain round of training, executing a layered federal migration learning strategy, and hierarchically aggregating received local models by an edge base station, wherein the specific method comprises the following steps: the edge base station aggregates basic characteristic layer models of all the devices to ensure the generalization performance of the models, and aggregates individual characteristic layer models of the non-migration devices to eliminate the influence of non-independent and same distributed data among the devices, and the formula is as follows:
Figure FDA0003301339630000038
Figure FDA0003301339630000041
wherein,
Figure FDA0003301339630000042
the base feature layer global model for the device accessing edge base station j, shared for all devices,
Figure FDA0003301339630000043
for the personality-level global model, for non-migration device sharing,
Figure FDA0003301339630000044
is a non-migrating device cluster accessing the edge base station j.
8. The method for generating an energy-efficient clustering federated edge learning strategy according to claim 1, wherein the formula of the system revenue function in step S5 is as follows:
Figure FDA0003301339630000045
where μ is a continuous variable and μ ∈ [0,1 ]]For adjusting the trade-off between learning performance and transmission power consumption, GmaxAnd EmaxThe highest accuracy and the maximum energy consumption of the system can be achieved.
9. The method for generating an energy-efficient clustering federation edge learning strategy according to claim 1, wherein in step S5, the edge access strategy is adjusted by deep reinforcement learning, and the specific process of the deep reinforcement learning is as follows:
s501, describing the edge correlation problem as a Markov process
Figure FDA0003301339630000046
The specific details are as follows:
(1) status of state
Figure FDA0003301339630000047
In the k-th round, the state is defined as S (k) { S ═ S1(k)，S2(k)，...，SN(k) Each item Si(k) Is defined as:
Si(k)＝{Ai(k-1)，βij(k)，Δi(k)}
wherein, Deltai(k) Indicates whether the learning accuracy is improved, i.e., Δ, compared to the k-1 roundi(k) 1 stands for improved accuracy, whereas Δi(k)＝0；
(2) Movement of
Figure FDA0003301339630000048
In the k-th round, the actions associate policies for the edge of each device:
A(k)＝{A1(k)，A2(k)，...，AN(k)}
each of whichItem Ai(k) Can be expressed as:
Ai(k)＝{aij(k)}
(3) reward
Figure FDA0003301339630000049
Set the reward as the objective function:
Figure FDA00033013396300000410
s502, selecting DQN as a basic frame, combining dulling DQN and double DQN to optimize an algorithm, using D3QN to solve the edge access problem, approximating a Q value function Q (S, A; theta) by a neural network with a parameter theta to represent the mapping relation between environment and action, and obtaining the output of the neural network through a Bellman equation:
Figure FDA0003301339630000051
wherein, S ', A ', theta ' are the state, action and corresponding parameters of the next time slot respectively;
two Q networks of the same structure but different parameters are used in DQN to improve the stability of the algorithm, one is the current Q network with the latest parameters to evaluate the current state-action cost function, the other is the target Q network with the past round parameters and keep the Q value unchanged for a period of time, the Q value of the current Q network is used as the input of the neural network, the goal of DQN is to minimize the difference between the two Q networks and define it as the loss function of DQN:
L(θ)＝E[(y-Q(S，A；θ))2]
s503, selecting the action corresponding to the maximum Q value in the current Q network by adopting a DDQN algorithm:
Figure FDA0003301339630000052
and then bringing the selected action into the target Q network to calculate a Q value:
y＝R(S，A)+γQ′(φ(S′)，Amax(S′；θ)；θ′)
s504, using blanking DQN to optimize the network structure, and dividing the network into two parts, namely a value function V (S, theta, alpha) only related to the state and a potential function A (S, A, theta, beta) related to both the state and the action, wherein theta is a common parameter of the two networks, alpha is a parameter unique to the value function, beta is a parameter unique to the potential function, and Q is the sum of the two functions:
Q(S，A，θ，α，β)＝V(S，θ，α)+A(S，A，θ，β)。
10. an energy-efficient clustered federated edge learning policy generation apparatus, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer processor, when executing the computer program, implements the energy-efficient clustered federated edge learning policy generation method of any one of claims 1-9.
CN202111191599.8A 2021-10-13 2021-10-13 Energy-efficient clustered federal edge learning strategy generation method and device Active CN113902021B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111191599.8A CN113902021B (en) 2021-10-13 2021-10-13 Energy-efficient clustered federal edge learning strategy generation method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111191599.8A CN113902021B (en) 2021-10-13 2021-10-13 Energy-efficient clustered federal edge learning strategy generation method and device

Publications (2)

Publication Number Publication Date
CN113902021A true CN113902021A (en) 2022-01-07
CN113902021B CN113902021B (en) 2024-06-21

Family

ID=79191774

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111191599.8A Active CN113902021B (en) 2021-10-13 2021-10-13 Energy-efficient clustered federal edge learning strategy generation method and device

Country Status (1)

Country Link
CN (1) CN113902021B (en)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114626306A (en) * 2022-03-22 2022-06-14 华北电力大学 Method and system for guaranteeing freshness of regulation and control information of park distributed energy
CN114666218A (en) * 2022-04-18 2022-06-24 中国科学技术大学苏州高等研究院 Efficient federal training method and device based on model migration
CN114912146A (en) * 2022-05-25 2022-08-16 中国信息通信研究院 Data information defense method and system under vertical federal architecture, electronic equipment and storage medium
CN114938372A (en) * 2022-05-20 2022-08-23 天津大学 Federal learning-based micro-grid group request dynamic migration scheduling method and device
CN115115021A (en) * 2022-01-17 2022-09-27 河南工业大学 Personalized federal learning method based on asynchronous updating of model parameters
CN115174412A (en) * 2022-08-22 2022-10-11 深圳市人工智能与机器人研究院 Dynamic bandwidth allocation method for heterogeneous federated learning system and related equipment
CN116209015A (en) * 2023-04-27 2023-06-02 合肥工业大学智能制造技术研究院 Edge network cache scheduling method, system and storage medium
CN116681126A (en) * 2023-06-06 2023-09-01 重庆邮电大学空间通信研究院 Asynchronous weighted federation learning method capable of adapting to waiting time
CN117076132A (en) * 2023-10-12 2023-11-17 北京邮电大学 Resource allocation and aggregation optimization method and device for hierarchical federal learning system
CN117373066A (en) * 2023-12-07 2024-01-09 华侨大学 Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method
CN117592580A (en) * 2023-11-21 2024-02-23 广东电网有限责任公司 Energy federation learning data selection method, device and energy federation learning system
CN117808128A (en) * 2024-02-29 2024-04-02 浪潮电子信息产业股份有限公司 Image processing method, federal learning method and device under heterogeneous data condition
WO2024087573A1 (en) * 2022-10-29 2024-05-02 华为技术有限公司 Federated learning method and apparatus

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190318268A1 (en) * 2018-04-13 2019-10-17 International Business Machines Corporation Distributed machine learning at edge nodes
CN110909865A (en) * 2019-11-18 2020-03-24 福州大学 Federated learning method based on hierarchical tensor decomposition in edge calculation
CN112804107A (en) * 2021-01-28 2021-05-14 南京邮电大学 Layered federal learning method for energy consumption adaptive control of equipment of Internet of things
WO2021169577A1 (en) * 2020-02-27 2021-09-02 山东大学 Wireless service traffic prediction method based on weighted federated learning
CN113467952A (en) * 2021-07-15 2021-10-01 北京邮电大学 Distributed federated learning collaborative computing method and system

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190318268A1 (en) * 2018-04-13 2019-10-17 International Business Machines Corporation Distributed machine learning at edge nodes
CN110909865A (en) * 2019-11-18 2020-03-24 福州大学 Federated learning method based on hierarchical tensor decomposition in edge calculation
WO2021169577A1 (en) * 2020-02-27 2021-09-02 山东大学 Wireless service traffic prediction method based on weighted federated learning
CN112804107A (en) * 2021-01-28 2021-05-14 南京邮电大学 Layered federal learning method for energy consumption adaptive control of equipment of Internet of things
CN113467952A (en) * 2021-07-15 2021-10-01 北京邮电大学 Distributed federated learning collaborative computing method and system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
吕洁娜;张家波;张祖凡;甘臣权;: "移动边缘计算卸载策略综述", 小型微型计算机系统, no. 09, 4 September 2020 (2020-09-04) *
张斌;李延晖;郭昊: "基于反向学习的跨种群差分进化算法", 计算机应用, vol. 37, no. 4, 10 April 2017 (2017-04-10) *

Cited By (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115115021A (en) * 2022-01-17 2022-09-27 河南工业大学 Personalized federal learning method based on asynchronous updating of model parameters
CN114626306B (en) * 2022-03-22 2023-01-24 华北电力大学 Method and system for guaranteeing freshness of regulation and control information of park distributed energy
CN114626306A (en) * 2022-03-22 2022-06-14 华北电力大学 Method and system for guaranteeing freshness of regulation and control information of park distributed energy
CN114666218A (en) * 2022-04-18 2022-06-24 中国科学技术大学苏州高等研究院 Efficient federal training method and device based on model migration
CN114666218B (en) * 2022-04-18 2023-11-10 中国科学技术大学苏州高等研究院 Efficient federal training method and device based on model migration
CN114938372A (en) * 2022-05-20 2022-08-23 天津大学 Federal learning-based micro-grid group request dynamic migration scheduling method and device
CN114938372B (en) * 2022-05-20 2023-04-18 天津大学 Federal learning-based micro-grid group request dynamic migration scheduling method and device
CN114912146A (en) * 2022-05-25 2022-08-16 中国信息通信研究院 Data information defense method and system under vertical federal architecture, electronic equipment and storage medium
CN115174412A (en) * 2022-08-22 2022-10-11 深圳市人工智能与机器人研究院 Dynamic bandwidth allocation method for heterogeneous federated learning system and related equipment
CN115174412B (en) * 2022-08-22 2024-04-12 深圳市人工智能与机器人研究院 Dynamic bandwidth allocation method for heterogeneous federal learning system and related equipment
WO2024087573A1 (en) * 2022-10-29 2024-05-02 华为技术有限公司 Federated learning method and apparatus
CN116209015A (en) * 2023-04-27 2023-06-02 合肥工业大学智能制造技术研究院 Edge network cache scheduling method, system and storage medium
CN116209015B (en) * 2023-04-27 2023-06-27 合肥工业大学智能制造技术研究院 Edge network cache scheduling method, system and storage medium
CN116681126A (en) * 2023-06-06 2023-09-01 重庆邮电大学空间通信研究院 Asynchronous weighted federation learning method capable of adapting to waiting time
CN116681126B (en) * 2023-06-06 2024-03-12 重庆邮电大学空间通信研究院 Asynchronous weighted federation learning method capable of adapting to waiting time
CN117076132B (en) * 2023-10-12 2024-01-05 北京邮电大学 Resource allocation and aggregation optimization method and device for hierarchical federal learning system
CN117076132A (en) * 2023-10-12 2023-11-17 北京邮电大学 Resource allocation and aggregation optimization method and device for hierarchical federal learning system
CN117592580A (en) * 2023-11-21 2024-02-23 广东电网有限责任公司 Energy federation learning data selection method, device and energy federation learning system
CN117592580B (en) * 2023-11-21 2024-10-08 广东电网有限责任公司 Energy federation learning data selection method, device and energy federation learning system
CN117373066A (en) * 2023-12-07 2024-01-09 华侨大学 Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method
CN117373066B (en) * 2023-12-07 2024-03-12 华侨大学 Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method
CN117808128A (en) * 2024-02-29 2024-04-02 浪潮电子信息产业股份有限公司 Image processing method, federal learning method and device under heterogeneous data condition
CN117808128B (en) * 2024-02-29 2024-05-28 浪潮电子信息产业股份有限公司 Image processing method and device under heterogeneous data condition

Also Published As

Publication number Publication date
CN113902021B (en) 2024-06-21

Similar Documents

Publication Publication Date Title
CN113902021A (en) High-energy-efficiency clustering federal edge learning strategy generation method and device
Bi et al. Lyapunov-guided deep reinforcement learning for stable online computation offloading in mobile-edge computing networks
Li et al. To talk or to work: Flexible communication compression for energy efficient federated learning over heterogeneous mobile edge devices
Wang et al. Computation offloading optimization for UAV-assisted mobile edge computing: a deep deterministic policy gradient approach
CN110968426B (en) Edge cloud collaborative k-means clustering model optimization method based on online learning
CN112598150B (en) Method for improving fire detection effect based on federal learning in intelligent power plant
CN113905347B (en) Cloud edge end cooperation method for air-ground integrated power Internet of things
CN112995913A (en) Unmanned aerial vehicle track, user association and resource allocation joint optimization method
Cui et al. Optimal rate adaption in federated learning with compressed communications
Lin et al. Contribution‐based Federated Learning client selection
Liu et al. Finch: Enhancing federated learning with hierarchical neural architecture search
CN116187483A (en) Model training method, device, apparatus, medium and program product
Song et al. Non-orthogonal multiple access assisted federated learning for UAV swarms: An approach of latency minimization
Jiang et al. Federated learning-based content popularity prediction in fog radio access networks
Chen et al. Joint data collection and resource allocation for distributed machine learning at the edge
Yu et al. Deep reinforcement learning for wireless networks
Liu et al. FedAGL: A communication-efficient federated vehicular network
Peng et al. Task offloading in multiple-services mobile edge computing: A deep reinforcement learning algorithm
Zhang et al. Device Scheduling and Assignment in Hierarchical Federated Learning for Internet of Things
Shi et al. Energy-efficient UAV-enabled computation offloading for industrial internet of things: a deep reinforcement learning approach
Liu et al. Group-based hierarchical federated learning: Convergence, group formation, and sampling
CN113965937A (en) Clustering federal learning-based content popularity prediction method in fog wireless access network
Cai et al. AutoInfer: Self-Driving Management for Resource-Efficient, SLO-Aware Machine= Learning Inference in GPU Clusters
CN116341679A (en) Design method of federal edge learning scheduling strategy with high aging
Chen et al. DDPG-based intelligent rechargeable fog computation offloading for IoT

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant