CN113177367A - High-energy-efficiency federal learning method and device, edge server and user equipment - Google Patents
High-energy-efficiency federal learning method and device, edge server and user equipment Download PDFInfo
- Publication number
- CN113177367A CN113177367A CN202110591514.9A CN202110591514A CN113177367A CN 113177367 A CN113177367 A CN 113177367A CN 202110591514 A CN202110591514 A CN 202110591514A CN 113177367 A CN113177367 A CN 113177367A
- Authority
- CN
- China
- Prior art keywords
- user equipment
- participating user
- participating
- local
- energy
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/12—Timing analysis or timing optimisation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- General Engineering & Computer Science (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Transfer Between Computers (AREA)
Abstract
The embodiment of the invention provides a high-energy-efficiency federal learning method, a high-energy-efficiency federal learning device, an edge server and user equipment, which are applied to the technical field of communication and comprise the following steps: determining the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment; issuing a global model and bandwidth information to each participating user equipment; receiving local model parameters uploaded by each participating user equipment; aggregating local model parameters uploaded by each participating user equipment to obtain an updated global model; judging whether the updated global model reaches the target precision; and if the updated global model does not reach the target precision, returning the participating user equipment determining the current communication turn and the bandwidth information to be distributed to each participating user equipment until the updated global model reaches the target precision. The method can realize federal learning in the edge network, and reduce learning time while ensuring model accuracy.
Description
Technical Field
The invention relates to the technical field of communication, in particular to a high-energy-efficiency federal learning method and device, an edge server and user equipment.
Background
With the growing concerns about data privacy and the increasing amount of data on end-user devices, traditional distributed machine learning, which requires users to share local data, has not been able to meet the demand. To address this problem, a new distributed machine learning framework, federal learning, has emerged. In a federal learning scene, training data are dispersed in a large number of user equipment, a server periodically broadcasts a global model (global model parameters) to the user equipment, the user equipment trains the model based on local data locally, and in uploading, a user does not need to return the local data but only needs to upload the model parameters obtained by training. And finally, aggregating the model parameters from each participating user equipment by the server to obtain model updating information, and repeatedly iterating the process until the model training reaches the target precision. Because the user does not need to upload local data in each communication turn, the training can ensure the privacy of the user, and meanwhile, the whole network flow load is reduced.
The task of federal learning is to enable the model training precision to reach the target precision, and in the process of federal learning, the server issues a global model to the user equipment, the user equipment performs model training based on local data, and the user equipment uploads model parameters obtained by local training to the server, and the like all need time, so that the model precision and the learning time are important indexes of federal learning.
Disclosure of Invention
The embodiment of the invention aims to provide a high-energy-efficiency federal learning method, a high-energy-efficiency federal learning device, an edge server and user equipment, so that the federal learning in an edge network is realized, the model precision is ensured, and the learning time is reduced. The specific technical scheme is as follows:
in a first aspect, an embodiment of the present invention provides an energy-efficient federal learning method, which is applied to an edge server in an edge network, and includes:
determining the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment; wherein the participating user equipment is user equipment participating in federal learning in the edge network;
issuing a global model and the bandwidth information to each participating user equipment so that each participating user equipment trains the global model based on local data to obtain local model parameters, and uploading the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information;
receiving local model parameters uploaded by each participating user equipment;
aggregating local model parameters uploaded by each participating user equipment to obtain an updated global model;
judging whether the updated global model reaches the target precision;
if the updated global model does not reach the target precision, returning the participating user equipment determining the current communication turn and the bandwidth information to be distributed to each participating user equipment, issuing the updated global model and the bandwidth information to each participating user equipment, and receiving the local model parameters uploaded by each participating user equipment; and aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model until the updated global model reaches the target precision.
Optionally, the determining the participating user equipment in the current communication turn and the bandwidth information to be allocated to each participating user equipment includes:
acquiring local calculation delay information, uploading delay information and energy information of each user equipment in the edge network;
determining a target optimization function based on the local calculation delay information, the uploading delay information and the energy information of each user equipment;
and solving the target optimization function to obtain the participating user equipment and the bandwidth information to be distributed to each participating user equipment.
Optionally, the determining a target optimization function based on the local computation delay information, the upload delay information, and the energy information of each user equipment includes:
according to the local calculation delay information and the uploading delay information of each user equipment, the formula is used
Determining learning completion time;
wherein D isrIndicates learning completion time of the r-th round, Sk(r) indicates whether user equipment k is selected as a sequence of participating user equipments in round r,
representing locally calculated delay information for user equipment k,
representing the uploading delay information of the user equipment k;
determining an objective optimization function according to the learning completion time and the energy information
Where B (r) denotes bandwidth information to be allocated to each participating user equipment, and B (r) ═ B1(r),...,BK(r)]T，S(r)＝[S1(r),...,SK(r)]T，Ak(r) indicates the round age of user equipment k, DrDenotes learning completion time, α denotes a predetermined constant, R denotes the total number of communication rounds, K denotes the total number of user equipments, ENk(r) represents energy information of the user equipment k in the r-th round, δ represents a preset energy threshold, Ek(r) indicates that user equipment k is in round rEnergy consumption of, BTkRepresenting the battery capacity of the user equipment k, ek(r) represents stored energy obtained based on the energy arrival conversion of the user equipment k in the r-th round,
indicating the energy arrival of user equipment k in round r.
Optionally, the solving the objective optimization function to obtain the participating user equipment and bandwidth information to be allocated to each participating user equipment includes:
converting the objective optimization function to:
wherein λ represents a predetermined constant, Qk(r) represents the virtual queue of energy in round r, Qk(r+1)＝max{Qk(r)+δ-ENk(r+1),0}，Qk(r +1) denotes the virtual queue of energy in round r +1, ENk(r +1) represents energy information of the user equipment k in the r +1 th round,
pkrepresenting the transmit power of the user equipment k, dmRepresenting the size of the data packet to be transmitted when uploading the local model parameters, hk(r) denotes the channel gain between the user equipment k and the edge server, N0Power spectral density, E, representing Gaussian noisek cRepresenting the training energy consumption of the user equipment k;
solving for S through multiple iterative optimizationk(r) and Bk(r)；
Wherein the one-time iterative optimization comprises the following steps:
Sk(r) after the fixation,
Sn/(r)＝0，
representing the set of user equipments selected for the current iteration, n/Representing divisions in a set of user equipment
A user equipment other than the device in (1);
based on Sn(r) converting the objective optimization function into
Wherein, Bn(r) is bandwidth information corresponding to the current iteration,
local calculation delay information representing the user equipment n selected by the current iteration;
b is to bek(r) fixing to obtain a converted objective optimization function
Wherein, ξ'1,k,rRepresenting the total time delay, ξ'2,k,rRepresenting the total energy consumption of the user equipment k in the r round based on the current bandwidth allocation strategy;
solving the converted target optimization function to obtain S corresponding to the current iterationk(r)。
Optionally, the converted objective optimization function is solved to obtain S corresponding to the current iterationk(r) comprising:
according to the formula- [ lambda alpha K delta + xi'2,k,rQk(r)-λAk(r)]Calculating a revenue ranking list of the selected user equipment; wherein the content of the first and second substances,
kmaxrepresenting the user equipment with the largest time delay;
adjusting the selected user equipment based on the profit ranking list, and when the currently selected user equipment reaches the target optimization function, taking a sequence formed by the currently selected user equipment as S corresponding to the current iterationk(r)。
In a second aspect, an embodiment of the present invention provides an energy-efficient federated learning method, which is applied to a participating user equipment, where the participating user equipment is a user equipment participating in federated learning in an edge network, and the method includes:
receiving a global model issued by an edge server and bandwidth information to be distributed to the participating user equipment;
training the global model based on the local data of the participating user equipment to obtain local model parameters;
and uploading the local model parameters to the edge server based on the bandwidth information so that the edge server receives the local model parameters uploaded by each participating user equipment, aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model, re-determining the participating user equipment and the bandwidth information to be distributed to each participating user equipment by the edge server when judging that the updated global model does not reach the target precision, and issuing the updated global model and the bandwidth information to each participating user equipment until the updated global model reaches the target precision.
The embodiment of the invention has the following beneficial effects:
according to the high-energy-efficiency federated learning method, the high-energy-efficiency federated learning device, the edge server and the user equipment, one-time communication turns comprise the steps that the edge server determines the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment, and issues the model and the bandwidth information to each participating user equipment, each participating user equipment trains a global model based on local data to obtain local model parameters, and uploads the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information. The edge server receives the local model parameters uploaded by each piece of participating user equipment, and aggregates the local model parameters uploaded by each piece of participating user equipment to obtain an updated global model. Meanwhile, when the communication round begins each time, the edge server determines the participating user equipment of the current communication round first, only the participating user equipment trains the global model based on local data to obtain local model parameters, and uploads the local model parameters to the edge server, and the edge server aggregates the local model parameters uploaded by each participating user equipment to obtain an updated global model.
In addition, in the embodiment of the invention, when the communication round begins each time, the edge server not only determines the participating user equipment of the current communication round, but also determines the bandwidth information to be distributed to each participating user equipment, and sends the bandwidth information to the participating user equipment, and the participating user equipment can upload the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information, so that the communication resources in the edge network can be fully utilized, and the influence on the federated learning caused by the limitation of the communication resources in the edge network is avoided.
Of course, not all of the advantages described above need to be achieved at the same time in the practice of any one product or method of the invention.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other embodiments can be obtained by referring to these drawings.
FIG. 1 is a system model diagram of an edge network;
FIG. 2 is a flow chart of a federated learning method that is applied to an edge server in an edge network in an embodiment of the present invention;
fig. 3 is a flowchart of determining participating user equipments in a current communication turn and bandwidth information to be allocated to each participating user equipment in the embodiment of the present invention;
FIG. 4 is a flow chart of a federated learning method as applied to a participating user device in an embodiment of the present invention;
fig. 5 is a schematic structural diagram of a federal learning apparatus applied to an edge server in an edge network in an embodiment of the present invention;
fig. 6 is a schematic structural diagram of a federal learning apparatus applied to a participating user equipment in an edge network in an embodiment of the present invention;
fig. 7 is a schematic structural diagram of an edge server according to an embodiment of the present invention;
fig. 8 is a schematic structural diagram of a user equipment according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived from the embodiments given herein by one of ordinary skill in the art, are within the scope of the invention.
In order to clearly illustrate the energy-efficient federal learning method provided by the embodiment of the present invention, a system model of an edge network is described below. As shown in fig. 1, the edge network includes an edge server and a plurality of user devices.
In federated learning, the learning objective of the server is to solve the optimal parameters to minimize the global loss function. A typical Federal learning System, training Process PackageThe method comprises the following steps: the server broadcasts an initial global model to each terminal device; the user equipment trains based on the local data set and the received global model, and updates respective local models; after training is finished, uploading the updated local model parameters to a server; the server aggregates all the received local model parameters and updates the global model; the above process iterates until the global model converges. Federal learning is applied to an edge network, and an edge server broadcasts a global model wgEach ue trains the global model based on its local data set to obtain a local model, e.g. the ue 1 trains to obtain the local model w1The user equipment 2 trains to obtain the local model w2The user equipment k trains to obtain the local model wkEach user equipment uploads the local model to the edge server, specifically, the local model can be represented by local model parameters, the edge server aggregates all received local model parameters, and updates the global model, and the above process is iterated until the global model converges, which can also be understood as meeting the target precision.
The system consists of an edge base station carrying a server as a federal learning server and K user equipment with energy collection capability. In a network, to
As a set of user equipment, the user equipment,
is the number of user equipments. For each user equipment
With a local data set
It has a size of DTk。
Is an input vector, yklIs input intoAnd outputting the corresponding output result of the vector. To be provided with
Representing the entire data set of the system. Total number of data samples
As shown in fig. 1, for each user equipment
Carrying a battery with a fixed capacity, the battery capacity being BTk，
Indicating the arrival of energy at the user equipment 1,
indicating the arrival of energy at the user equipment 2,
representing the energy arrival of the user equipment k.
In order to better complete federal learning, in the embodiment of the invention, the edge server formulates a user selection and communication resource allocation strategy because the computing resources and network communication resources of the user equipment are limited in a wireless edge scene.
For federal learning, the most important indicators are model accuracy and training completion time. Meanwhile, when the federal study is applied to an important application scene and a wireless edge network, the optimization aiming at the indexes needs to be combined with the characteristics of the network: communication resources, end user computing resources, and terminal energy resources in the network are all limited. Therefore, research on user selection and resource allocation strategies for user equipment participating in learning is often required.
Considering that for user equipment with limited energy resources in the edge network, the energy consumption caused by the training and uploading tasks of the federal learning model is a big challenge for the survival time of the federal learning network. Energy collection from natural environments such as solar energy, wind energy and the like is considered to be an effective method for prolonging the survival time of energy-limited equipment, and a new thought is provided for guaranteeing the federal learning effect. Therefore, in the federal learning network, based on the energy harvesting technology, the intensive research aiming at the problems of training effect and efficiency, and stability of energy in the network, namely, survival time of the network, has an intensive significance for the application of the federal learning in the edge network.
The energy-efficient federal learning method provided by the embodiment of the invention is explained in detail below.
The embodiment of the invention provides a high-energy-efficiency federal learning method, which is applied to an edge server in an edge network and can comprise the following steps:
determining the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment; the participating user equipment is user equipment participating in federal learning in an edge network;
issuing a global model and bandwidth information to each participating user equipment so that each participating user equipment trains the global model based on local data to obtain local model parameters, and uploading the local model parameters obtained by local training of the participating user equipment to an edge server based on the bandwidth information;
receiving local model parameters uploaded by each participating user equipment;
aggregating local model parameters uploaded by each participating user equipment to obtain an updated global model;
judging whether the updated global model reaches the target precision;
if the updated global model does not reach the target precision, returning the participating user equipment determining the current communication turn and the bandwidth information to be distributed to each participating user equipment, issuing the updated global model and the bandwidth information to each participating user equipment, and receiving the local model parameters uploaded by each participating user equipment; and aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model until the updated global model reaches the target precision.
In the embodiment of the invention, the primary communication turn comprises the steps that the edge server determines the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment, and issues the model and the bandwidth information to each participating user equipment, each participating user equipment trains the global model based on local data to obtain local model parameters, and uploads the local model parameters obtained by the local training of the participating user equipment to the edge server based on the bandwidth information. The edge server receives the local model parameters uploaded by each piece of participating user equipment, and aggregates the local model parameters uploaded by each piece of participating user equipment to obtain an updated global model. Meanwhile, when the communication round begins each time, the edge server determines the participating user equipment of the current communication round first, only the participating user equipment trains the global model based on local data to obtain local model parameters, and uploads the local model parameters to the edge server, and the edge server aggregates the local model parameters uploaded by each participating user equipment to obtain an updated global model.
In addition, in the embodiment of the invention, when the communication round begins each time, the edge server not only determines the participating user equipment of the current communication round, but also determines the bandwidth information to be distributed to each participating user equipment, and sends the bandwidth information to the participating user equipment, and the participating user equipment can upload the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information, so that the communication resources in the edge network can be fully utilized, and the influence on the federated learning caused by the limitation of the communication resources in the edge network is avoided.
Fig. 2 is a flowchart of a federal learning method applied to an edge server in an edge network in an embodiment of the present invention, and referring to fig. 2, the federal learning method applied to an edge server in an edge network in an embodiment of the present invention may include:
s201, determining the participating user equipment of the current communication turn and the bandwidth information to be allocated to each participating user equipment.
The participating user equipment is the user equipment participating in federal learning in the edge network.
And for each participating user equipment, the bandwidth information is used for instructing the participating user equipment to upload local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information.
In an alternative embodiment, as shown in fig. 3, S201 may include:
s2011, local computation delay information, upload delay information, and energy information of each ue in the edge network are obtained.
And S2012, determining a target optimization function based on the local calculation delay information, the uploading delay information and the energy information of each user equipment.
According to the local calculation delay information and the uploading delay information of each user equipment, the formula is used
Determining learning completion time;
wherein D isrIndicates learning completion time of the r-th round, Sk(r) indicates whether user equipment k is selected as a sequence of participating user equipments in round r,
representing locally calculated delay information for user equipment k,
representing the uploading delay information of the user equipment k;
determining a target optimization function according to learning completion time and energy information
The objective optimization function may be referred to as equation 1. Where B (r) denotes bandwidth information to be allocated to each participating user equipment, and B (r) ═ B1(r),...,BK(r)]T，S(r)＝[S1(r),...,SK(r)]T，Ak(r) indicates the round age of user equipment k, DrDenotes learning completion time, α denotes a predetermined constant for adjusting a trade-off relationship between round age and learning time, R denotes a total number of rounds of communication, K denotes a total number of user equipments, ENk(r) represents energy information of the user equipment k in the r-th round, δ represents a preset energy threshold, Ek(r) represents the energy consumption of user equipment k in round r, BTkRepresenting the battery capacity of the user equipment k, ek(r) represents stored energy obtained based on the energy arrival conversion of the user equipment k in the r-th round,
indicating the energy arrival of user equipment k in round r.
The iterative process of each federal learning is referred to as a communication turn, i.e., a global communication turn. According to the federal learning completion process, the completion time of each communication turn is composed of global model aggregation and broadcasting time, local model training time and local model uploading time. The server is considered to have a strong computing power and transmission power and to allocate the full bandwidth when broadcasting the global model, so the global model aggregation and the broadcast time are negligible. In the embodiment of the present invention, a Frequency Division Multiple Access (FDMA) system may be adopted. Meanwhile, the local model training of each user equipment participating in the federal learning is started locally after receiving the global model. Thus, the completion time of each communication round is determined by all books participating in the training of the communication roundThe user who trains and uploads the model for the longest time. Learning completion time D of the r-th round as shown in the following formularCan be expressed as follows:
wherein S isk(r) is a variable from 0 to 1, when SkIf (r) is 1, it means that the ue k is selected to participate in training in the r-th round, i.e. to participate in federal learning, SkIf (r) ═ 0, then it means that the ue k is not selected to participate in the training in the r-th round.
The calculated delay, i.e. the locally calculated delay information,
the upload delay information is the upload delay of the user equipment k.
Let k be the number of cycles per second of a Central Processing Unit (CPU) of the UEkThe parameter is indicative of the computational power of the user device. CkThe number of CPU cycles, I, required to compute a sample of data for user equipment kkThe local iteration times in each training of the user equipment k, the local calculation time delay of the user equipment k
Can be expressed as follows:
wherein, DTkRepresenting the local data set size.
In a frequency division multiple access system, all user equipments and edge servers communicate over a common spectrum sharing a total bandwidth B. With Bk(r) represents the bandwidth allocated to user equipment k in communication round r. Data rate of transmission rate of user equipment k
Can be expressed as follows:
wherein p iskRepresenting the transmit power, h, of the user equipment kk(r) denotes the channel gain between the user equipment k and the edge server, N0Representing the power spectral density of gaussian noise. Wherein the channel state is assumed to remain unchanged in one transmission.
According to the above formula, the upload delay of the user equipment k can be expressed as follows:
wherein d ismAnd the size of a data packet required to be transmitted when the local model parameters are uploaded is represented.
In federal learning, since training data is stored at individual user devices in a distributed manner, energy stability of the user devices is very important for sustainable learning. Meanwhile, in the wireless edge network, due to the limited communication resources, the user equipment participating in federal learning needs to be scheduled and resource allocated, so that the learning task can be completed more quickly and better. The invention aims to research a user selection and resource allocation combined strategy aiming at optimizing the federal learning time and accuracy while maintaining the energy stability of a system in a federal learning network participated by users with energy collection capability.
The energy stability of the user equipment in the embodiment of the present invention is explained below.
For each user equipment, energy consumption comes from local model training and model parameter uploading. With immediate energy consumption Ek(r) is represented as follows:
the energy consumption for local model training at the user equipment depends on the training data size and complexity, the CPU architecture and training capabilities of the user equipment. The training energy consumption of user equipment k can be expressed as follows:
where κ denotes the effective switched capacitance, this value depending on the chip structure, DTkRepresenting the local data set size.
Energy consumption for uploading at user equipment k
Determined by the transmission time and the transmission power, there are:
to support long term learning, user devices equipped with energy harvesting components harvest energy from the environment, such as solar and wind energy, at the beginning of each round of communication. The user equipment will acquire the energy and store it in the battery. It is assumed that the energy acquired by the user equipment is independent and co-distributed in different communication rounds. By using
Indicating the energy arrival of user equipment k in round r. After each energy harvest, the user device may store the newly harvested energy at a conversion rate. Denote the stored energy of the user equipment k in the r-th round as ek(r), there may be:
a representation of the energy level of user k at the start of the r-th round of global training at user k can thus be obtained as follows:
ENk(r+1)＝min{ENk(r)-Ek(r)+ek(r),BTk}
wherein, BTkIs the battery capacity of the user equipment k.
It is assumed that data samples of different user equipments, i.e. local data distribution of the user equipments, are not independently and equally distributed. Thus, the absence of any user equipment affects learning performance. To avoid that local model parameters, i.e. data loss, cannot be contributed due to energy exhaustion or long term low battery of the user equipment, a certain level of energy should be stored in the battery of each user equipment for future use, the stability constraint of the battery energy level is given by:
where δ is a predetermined energy level threshold, i.e. a predetermined energy threshold.
In order to optimize the effect of federal learning, the round age of the user equipment is optimized to represent that the user participates in global aggregation twice in a neighborhood, namely the round difference between federal learning. The engagement of individual users is guaranteed by minimizing the round age of the user equipment. Round age of user AkThe evolution law of (r) can be expressed as follows:
Ak(r+1)＝[Ak(r)+1][1-Sk(r)]
wherein A isk(r) represents the round age of user equipment k in round r, Ak(r +1) round age of user equipment k in round r + 1.
The method aims to jointly optimize the effect of federal learning and learning completion time while maintaining the energy stability of user equipment. The optimization problem can be expressed as an objective optimization function as shown in equation 1 above.
S2013, solving the objective optimization function to obtain the participating user equipment and the bandwidth information to be distributed to each participating user equipment.
Converting the objective optimization function to:
the transformed objective optimization function may be referred to as equation 2. Wherein λ represents a predetermined constant, Qk(r) represents the virtual queue of energy in round r, Qk(r+1)＝max{Qk(r)+δ-ENk(r+1),0}，Qk(r +1) denotes the virtual queue of energy in round r +1, ENk(r +1) represents energy information of the user equipment k in the r +1 th round,
pkrepresenting the transmit power of the user equipment k, dmRepresenting the size of the data packet to be transmitted when uploading the local model parameters, hk(r) denotes the channel gain between the user equipment k and the edge server, N0Represents the power spectral density of the gaussian noise,
representing the training energy consumption of the user equipment k.
In order to decouple long-term energy limitation, based on the Lyapunov optimization theory, an energy virtual queue Q can be constructedk(r), the dynamics of which are as follows:
Qk(r+1)＝max{Qk(r)+δ-ENk(r+1),0}
further, the lyapunov equation is established as follows:
the lyapunov Drift-plus-penalty equation is constructed to measure the tradeoff between the energy stability constraint and the optimization objective, and can be expressed as follows:
wherein the content of the first and second substances,
To solve the minimization above, the upper bound can be minimized, as follows:
the substitution can be obtained:
wherein the content of the first and second substances,
is a constant value, and is characterized in that,
r is constant for the round.
Let a be a1+a2Then, there are:
substitution into
It is possible to obtain:
wherein
Are constants related to k, r only.
That is, the objective optimization function transforms the objective optimization function shown in equation 2 above.
Solving for S through multiple iterative optimizationk(r) and Bk(r)。
Wherein the one-time iterative optimization comprises the following steps:
Sk(r) after the fixation,
Sn/(r)＝0，
representing the set of user equipments selected for the current iteration, n/Representing divisions in a set of user equipment
A device other than the device in (1); that is to say SkAfter (r) is fixed, for
Is a subset of
User equipment n in (1), having Sn(r)＝1，
Is 0, i.e. for the division in the user equipment set
User equipment other than the device in (1), Sn/(r)＝0。
Based on Sn(r) converting the objective optimization function into
Wherein, Bn(r) is bandwidth information corresponding to the current iteration,
local calculation delay information representing the user equipment n selected by the current iteration;
b is to bek(r) fixing to obtain a converted objective optimization function
Wherein ξ1',k,rRepresenting the total delay, ξ, of user equipment k in round r based on the current bandwidth allocation policy2',k,rRepresenting the total energy consumption of the user equipment k in the r round based on the current bandwidth allocation strategy;
solving the converted target optimization function to obtain S corresponding to the current iterationk(r)。
In particular, it can be according to the formula- [ λ α K δ + ξ2',k,rQk(r)-λAk(r)]Calculating a revenue ranking list of the selected user equipment; wherein the content of the first and second substances,
kmaxrepresenting the user equipment with the largest time delay;
adjusting the selected user equipment based on the profit ranking list, and when the currently selected user equipment reaches the target optimization function, taking the sequence formed by the currently selected user equipment as the S corresponding to the current iterationk(r)。
In order to reduce the complexity of problem solution, an iteration mode is adopted for Sk(r) and Bk(r) are solved separately.
When S isk(r) fixed, set the selected user set as
Then there is
The objective function can be converted into:
wherein
λ α K is equal to Bn(r) independent constants, so the above objective function is equivalent to:
this problem is easily confirmed as a convex problem.
When B is presentk(r) fixed, the objective function can be converted to:
wherein
Irrespective of Sk(r) an independent constant a, the optimization problem being equivalent to
Considering the benefits of selecting a user as
-[λαKδ+ξ′2,k,rQk(r)-λAk(r)]
Wherein
According to the above formula, a revenue ranking list of the selected users can be obtained, namely, the ranking of the user selection priority is obtained.
In each communication turn, the edge server jointly optimizes the user selection and bandwidth allocation strategy based on the information of the turn age, the energy virtual queue, the energy consumption, the learning delay and the uploading delay of the user equipment. In general, users with older round of ages will be preferentially selected to optimize learning accuracy, users with larger energy virtual queue value and smaller energy consumption will be preferentially selected to maintain system energy stability, and users with smaller training and uploading delay will be preferentially selected to optimize learning completion time.
S202, issuing the global model and the bandwidth information to each participating user equipment to enable each participating user equipment to train the global model based on local data to obtain local model parameters, and uploading the local model parameters obtained by local training of the participating user equipment to an edge server based on the bandwidth information.
S203, receiving the local model parameters uploaded by each participating user equipment.
And S204, aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model.
And S205, judging whether the updated global model reaches the target precision.
The target accuracy can be determined according to actual requirements. Specifically, whether the model parameters converge may also be understood by determining whether the model parameters can achieve minimization of the global loss function.
If the updated global model does not reach the target precision, the process returns to S201. And if the updated global model reaches the target precision, ending the whole process.
Entering the next communication turn, determining the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment, issuing updated global model and bandwidth information to each participating user equipment, and receiving local model parameters uploaded by each participating user equipment; and aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model, and performing multiple communication rounds until the updated global model reaches the target precision.
The primary communication turn comprises the steps that the edge server determines the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment, and sends the model and the bandwidth information to each participating user equipment, each participating user equipment trains the global model based on local data to obtain local model parameters, and uploads the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information. The edge server receives the local model parameters uploaded by each piece of participating user equipment, and aggregates the local model parameters uploaded by each piece of participating user equipment to obtain an updated global model. Meanwhile, when the communication round begins each time, the edge server determines the participating user equipment of the current communication round first, only the participating user equipment trains the global model based on local data to obtain local model parameters, and uploads the local model parameters to the edge server, and the edge server aggregates the local model parameters uploaded by each participating user equipment to obtain an updated global model.
And when the communication round begins each time, the edge server not only determines the participating user equipment of the current communication round, but also determines the bandwidth information to be distributed to each participating user equipment, and transmits the bandwidth information to the participating user equipment, and the participating user equipment can upload the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information, so that the communication resources in the edge network can be fully utilized, and the influence on the federated learning caused by the limitation of the communication resources in the edge network is avoided.
In the embodiment of the invention, when each communication turn starts, the edge server solves the user selection strategy and the bandwidth allocation strategy according to the iterative algorithm, namely determines the participating user equipment of the current communication turn and the bandwidth information to be allocated to each participating user equipment. In an implementation manner, a user selection policy may be initialized first, a bandwidth allocation policy is determined, then user screening is performed according to the determined user priority ranking, and the above steps are repeated until convergence. And determining a user selection and bandwidth allocation combined strategy, broadcasting the aggregated global model parameters, and starting training and uploading of each selected terminal device. And iterating until the federal learning converges, and understanding that the model reaches the target precision.
Fig. 4 is a flowchart of a federated learning method applied to a participating user equipment in an embodiment of the present invention, where the participating user equipment is a user equipment participating in federated learning in an edge network. Referring to fig. 4, the federal learning applied to the participating user equipment according to the embodiment of the present invention may include:
s401, receiving a global model issued by an edge server and bandwidth information to be distributed to participating user equipment;
s402, training the global model based on the local data of the participating user equipment to obtain local model parameters;
and S403, uploading local model parameters to the edge server based on the bandwidth information, so that the edge server receives the local model parameters uploaded by each participating user equipment, aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model, when the updated global model is judged not to reach the target precision, the edge server re-determines the participating user equipment and the bandwidth information to be distributed to each participating user equipment, and issues the updated global model and the bandwidth information to each participating user equipment until the updated global model reaches the target precision.
Corresponding to the federate learning method applied to the edge server in the edge network provided in the foregoing embodiment, an embodiment of the present invention further provides an energy-efficient federate learning apparatus, which is applied to the edge server in the edge network, as shown in fig. 5, and includes:
a determining module 501, configured to determine participating user equipment in a current communication turn and bandwidth information to be allocated to each participating user equipment; the participating user equipment is user equipment participating in federal learning in an edge network;
an issuing module 502, configured to issue a global model and bandwidth information to each participating user equipment, so that each participating user equipment trains the global model based on local data to obtain a local model parameter, and uploads the local model parameter obtained by local training of the participating user equipment to an edge server based on the bandwidth information;
a receiving module 503, configured to receive local model parameters uploaded by each participating user equipment;
an aggregation module 504, configured to aggregate local model parameters uploaded by each participating user equipment to obtain an updated global model;
a judging module 505, configured to judge whether the updated global model reaches the target precision;
if the updated global model does not reach the target precision, the process returns to the determining module 501, and the determining module 501, the issuing module 502, the receiving module 503, the aggregating module 504 and the judging module 505 are executed again in sequence until the updated global model reaches the target precision.
Optionally, the determining module 501 is specifically configured to obtain local computation delay information, upload delay information, and energy information of each user equipment in the edge network; determining a target optimization function based on the local calculation delay information, the uploading delay information and the energy information of each user equipment; and solving the objective optimization function to obtain the participating user equipment and the bandwidth information to be distributed to each participating user equipment.
Optionally, the determining module 501 is specifically configured to calculate delay information and upload delay information locally according to each ue, and use a formula to determine the delay information
Determining learning completion time;
wherein D isrIndicates learning completion time of the r-th round, Sk(r) indicates whether user equipment k is selected as a sequence of participating user equipments in round r,
representing locally calculated delay information for user equipment k,
representing the uploading delay information of the user equipment k;
determining a target optimization function according to learning completion time and energy information
Where B (r) denotes bandwidth information to be allocated to each participating user equipment, and B (r) ═ B1(r),...,BK(r)]T，S(r)＝[S1(r),...,SK(r)]T，Ak(r) indicates the round age of user equipment k, DrDenotes learning completion time, α denotes a predetermined constant, R denotes the total number of communication rounds, K denotes the total number of user equipments, ENk(r) represents energy information of the user equipment k in the r-th round, δ represents a preset energy threshold, Ek(r) represents the energy consumption of user equipment k in round r, BTkRepresenting the battery capacity of the user equipment k, ek(r) represents stored energy obtained based on the energy arrival conversion of the user equipment k in the r-th round,
indicating the energy arrival of user equipment k in round r.
Optionally, the determining module 501 is specifically configured to convert the objective optimization function into:
wherein λ represents a predetermined constant, Qk(r) represents the virtual queue of energy in round r, Qk(r+1)＝max{Qk(r)+δ-ENk(r+1),0}，Qk(r +1) denotes the virtual queue of energy in round r +1, ENk(r +1) represents energy information of the user equipment k in the r +1 th round,
pkrepresenting the transmit power of the user equipment k, dmRepresenting the size of the data packet to be transmitted when uploading the local model parameters, hk(r) denotes the channel gain between the user equipment k and the edge server, N0Represents the power spectral density of the gaussian noise,
representing the training energy consumption of the user equipment k;
solving for S through multiple iterative optimizationk(r) and Bk(r)；
Wherein the one-time iterative optimization comprises the following steps:
Sk(r) after the fixation,
Sn/(r)＝0，
representing the set of user equipments selected for the current iteration, n/Representing a set of user equipmentsHezhongwei
A user equipment other than the device in (1);
based on Sn(r) converting the objective optimization function into
Wherein, Bn(r) is bandwidth information corresponding to the current iteration,
local calculation delay information representing the user equipment n selected by the current iteration;
b is to bek(r) fixing to obtain a converted objective optimization function
Wherein, ξ'1,k,rRepresenting the total time delay, ξ'2,k,rRepresenting the total energy consumption of the user equipment k in the r round based on the current bandwidth allocation strategy;
solving the converted target optimization function to obtain S corresponding to the current iterationk(r)。
Optionally, the determining module 501 is specifically configured to determine according to a formula- [ λ α K δ + ξ'2,k,rQk(r)-λAk(r)]Calculating a revenue ranking list of the selected user equipment; wherein the content of the first and second substances,
kmaxrepresenting the user equipment with the largest time delay;
adjusting the selected user equipment based on the profit ranking list, and when the currently selected user equipment reaches the target optimization function, taking the sequence formed by the currently selected user equipment as the S corresponding to the current iterationk(r)。
The federate learning device applied to the edge server in the edge network provided by the embodiment of the invention is a device applying the federate learning method applied to the edge server in the edge network, so that all embodiments of the federate learning method applied to the edge server in the edge network are suitable for the device and can achieve the same or similar beneficial effects.
Corresponding to the federate learning method applied to the user equipment participating in the federate learning in the edge network provided in the above embodiment, an embodiment of the present invention further provides a federate learning apparatus applied to the user equipment participating in the federate learning in the edge network, where the user equipment participating in the federate learning in the edge network is, as shown in fig. 6, the method includes:
a receiving module 601, configured to receive a global model issued by an edge server and bandwidth information to be allocated to a participating user equipment;
a training module 602, configured to train the global model based on local data of participating user equipment to obtain local model parameters;
the uploading module 603 is configured to upload local model parameters to the edge server based on the bandwidth information, so that the edge server receives the local model parameters uploaded by each participating user equipment, aggregates the local model parameters uploaded by each participating user equipment to obtain an updated global model, and when it is determined that the updated global model does not reach the target accuracy, the edge server re-determines the participating user equipment and the bandwidth information to be allocated to each participating user equipment, and issues the updated global model and the bandwidth information to each participating user equipment until the updated global model reaches the target accuracy.
The federate learning device applied to the participating user equipment in the edge network provided by the embodiment of the invention is a device applying the federate learning method applied to the participating user equipment in the edge network, so that all embodiments of the federate learning method applied to the participating user equipment in the edge network are applicable to the device and can achieve the same or similar beneficial effects.
An embodiment of the present invention further provides an edge server in an edge network, as shown in fig. 7, including a processor 701, a communication interface 702, a memory 703 and a communication bus 704, where the processor 701, the communication interface 702, and the memory 703 complete mutual communication through the communication bus 704.
A memory 703 for storing a computer program;
the processor 701 is configured to implement the method steps of the federal learning method applied to an edge server in an edge network when executing the program stored in the memory 703.
An embodiment of the present invention further provides a user equipment in an edge network, as shown in fig. 8, including a processor 801, a communication interface 802, a memory 803, and a communication bus 804, where the processor 801, the communication interface 802, and the memory 803 complete mutual communication through the communication bus 804.
A memory 803 for storing a computer program;
the processor 801 is configured to implement the method steps of the federal learning method applied to the ue in the edge network when executing the program stored in the memory 803.
The communication bus mentioned in the above electronic devices (such as the edge server and the user equipment) may be a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The communication bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is shown, but this does not mean that there is only one bus or one type of bus.
The communication interface is used for communication between the electronic equipment and other equipment.
The Memory may include a Random Access Memory (RAM) or a Non-Volatile Memory (NVM), such as at least one disk Memory. Optionally, the memory may also be at least one memory device located remotely from the processor.
The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), and the like; but also Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components.
In yet another embodiment of the present invention, a computer-readable storage medium is further provided, in which a computer program is stored, and the computer program, when executed by a processor, implements the method steps of the above federated learning method applied to user equipment in an edge network.
In yet another embodiment provided by the present invention, a computer-readable storage medium is further provided, in which a computer program is stored, which, when being executed by a processor, implements the above method steps of the federated learning method applied to an edge server in an edge network.
In yet another embodiment provided by the present invention, there is also provided a computer program product containing instructions which, when run on a computer, cause the computer to perform the above-described method steps of the federated learning method as applied to an edge server in an edge network.
In yet another embodiment provided by the present invention, there is also provided a computer program product containing instructions which, when run on a computer, cause the computer to perform the above-described method steps of the federal learning method applied to user equipment in an edge network.
In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, cause the processes or functions described in accordance with the embodiments of the invention to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, from one website site, computer, server, or data center to another website site, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that incorporates one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
All the embodiments in the present specification are described in a related manner, and the same and similar parts among the embodiments may be referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, as for the apparatus, the edge server, the user equipment, the computer-readable storage medium, and the computer program product embodiment, since they are substantially similar to the method embodiment, the description is relatively simple, and it is sufficient to refer to the partial description of the method embodiment for relevant points.
The above description is only for the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.
Claims (10)
1. An energy-efficient federated learning method applied to an edge server in an edge network, comprising:
determining the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment; wherein the participating user equipment is user equipment participating in federal learning in the edge network;
issuing a global model and the bandwidth information to each participating user equipment so that each participating user equipment trains the global model based on local data to obtain local model parameters, and uploading the local model parameters obtained by local training of the participating user equipment to the edge server based on the bandwidth information;
receiving local model parameters uploaded by each participating user equipment;
aggregating local model parameters uploaded by each participating user equipment to obtain an updated global model;
judging whether the updated global model reaches the target precision;
if the updated global model does not reach the target precision, returning the participating user equipment determining the current communication turn and the bandwidth information to be distributed to each participating user equipment, issuing the updated global model and the bandwidth information to each participating user equipment, and receiving the local model parameters uploaded by each participating user equipment; and aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model until the updated global model reaches the target precision.
2. The method of claim 1, wherein determining the participating user devices of the current communication turn and bandwidth information to be allocated to each participating user device comprises:
acquiring local calculation delay information, uploading delay information and energy information of each user equipment in the edge network;
determining a target optimization function based on the local calculation delay information, the uploading delay information and the energy information of each user equipment;
and solving the target optimization function to obtain the participating user equipment and the bandwidth information to be distributed to each participating user equipment.
3. The method of claim 2, wherein determining the objective optimization function based on the locally computed delay information, the uploaded delay information, and the energy information of each ue comprises:
according to the local calculation delay information and the uploading delay information of each user equipment, the formula is used
Determining learning completion time;
wherein D isrIndicates learning completion time of the r-th round, Sk(r) indicates whether user equipment k is selected as a sequence of participating user equipments in round r,
representing locally calculated delay information for user equipment k,
representing the uploading delay information of the user equipment k;
determining an objective optimization function according to the learning completion time and the energy information
Where B (r) denotes bandwidth information to be allocated to each participating user equipment, and B (r) ═ B1(r),...,BK(r)]T，S(r)＝[S1(r),...,SK(r)]T，Ak(R) represents a round age of the user equipment K, α represents a predetermined constant, R represents a total number of rounds of communication rounds, K represents a total number of user equipments,
representing a set of user equipments, ENk(r) represents energy information of the user equipment k in the r-th round, δ represents a preset energy threshold, Ek(r) represents the energy consumption of user equipment k in round r, BTkRepresenting the battery capacity of the user equipment k, ek(r) represents stored energy obtained based on the energy arrival conversion of the user equipment k in the r-th round,
indicating the energy arrival of user equipment k in round r.
4. The method of claim 3, wherein solving the objective optimization function to obtain the participating user devices and bandwidth information to be allocated to each participating user device comprises:
converting the objective optimization function to:
wherein λ represents a predetermined constant, Qk(r) represents the virtual queue of energy in round r, Qk(r+1)＝max{Qk(r)+δ-ENk(r+1),0}，Qk(r +1) denotes the virtual queue of energy in round r +1, ENk(r +1) represents energy information of the user equipment k in the r +1 th round,
pkrepresenting the transmit power of the user equipment k, dkRepresenting the size of the data packet to be transmitted when uploading the local model parameters, hk(r) denotes the channel gain between the user equipment k and the edge server, N0Represents the power spectral density of the gaussian noise,
representing the training energy consumption of the user equipment k;
solving for S through multiple iterative optimizationk(r) and Bk(r)；
Wherein the one-time iterative optimization comprises the following steps:
Sk(r) after fixation, Sn(r)＝1,
Sn/(r)＝0，
Representing the set of user equipments selected for the current iteration, n/Representing divisions in a set of user equipment
A user equipment other than the device in (1);
based on Sn(r) converting the objective optimization function into
Wherein, Bn(r) is bandwidth information corresponding to the current iteration,
local calculation delay information representing the user equipment n selected by the current iteration;
b is to bek(r) fixing to obtain a converted objective optimization function
Wherein, ξ'1,k,rRepresenting the total time delay, ξ'2,k,rRepresenting the total energy consumption of the user equipment k in the r round based on the current bandwidth allocation strategy;
solving the converted target optimization function to obtain S corresponding to the current iterationk(r)。
5. The method of claim 4, wherein the transformed objective optimization function is solved to obtain S corresponding to the current iterationk(r) comprising:
according to the formula- [ lambda alpha K delta + xi'2,k,rQk(r)-λAk(r)]Calculating a revenue ranking list of the selected user equipment; wherein the content of the first and second substances,
kmaxrepresenting the user equipment with the largest time delay;
adjusting the selected user equipment based on the profit ranking list, and when the currently selected user equipment reaches the target optimization function, taking a sequence formed by the currently selected user equipment as S corresponding to the current iterationk(r)。
6. An energy-efficient federated learning method is applied to a participating user equipment, wherein the participating user equipment is a user equipment participating in federated learning in an edge network, and the method comprises the following steps:
receiving a global model issued by an edge server and bandwidth information to be distributed to the participating user equipment;
training the global model based on the local data of the participating user equipment to obtain local model parameters;
and uploading the local model parameters to the edge server based on the bandwidth information so that the edge server receives the local model parameters uploaded by each participating user equipment, aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model, re-determining the participating user equipment and the bandwidth information to be distributed to each participating user equipment by the edge server when judging that the updated global model does not reach the target precision, and issuing the updated global model and the bandwidth information to each participating user equipment until the updated global model reaches the target precision.
7. The utility model provides a bang learning device which characterized in that, is applied to the edge server in the edge network, includes:
the determining module is used for determining the participating user equipment of the current communication turn and the bandwidth information to be distributed to each participating user equipment; wherein the participating user equipment is user equipment participating in federal learning in the edge network;
the issuing module is used for issuing a global model and the bandwidth information to each participating user equipment so that each participating user equipment trains the global model based on local data to obtain local model parameters, and uploads the local model parameters obtained by the participating user equipment in local training to the edge server based on the bandwidth information;
the receiving module is used for receiving the local model parameters uploaded by each piece of participating user equipment;
the aggregation module is used for aggregating the local model parameters uploaded by each participating user equipment to obtain an updated global model;
the judging module is used for judging whether the updated global model reaches the target precision;
and if the updated global model does not reach the target precision, returning to the determining module, and executing the determining module, the issuing module, the receiving module, the aggregation module and the judging module again in sequence until the updated global model reaches the target precision.
8. The apparatus according to claim 7, wherein the determining module is specifically configured to obtain local computation delay information, upload delay information, and energy information of each user equipment in the edge network; determining a target optimization function based on the local calculation delay information, the uploading delay information and the energy information of each user equipment; and solving the target optimization function to obtain the participating user equipment and the bandwidth information to be distributed to each participating user equipment.
9. An edge server in an edge network is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus;
a memory for storing a computer program;
a processor for implementing the method steps of any one of claims 1 to 5 when executing a program stored in the memory.
10. The user equipment in the edge network is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus;
a memory for storing a computer program;
a processor for implementing the method steps of claim 6 when executing a program stored in the memory.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110591514.9A CN113177367B (en) | 2021-05-28 | 2021-05-28 | High-energy-efficiency federal learning method and device, edge server and user equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110591514.9A CN113177367B (en) | 2021-05-28 | 2021-05-28 | High-energy-efficiency federal learning method and device, edge server and user equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113177367A true CN113177367A (en) | 2021-07-27 |
| CN113177367B CN113177367B (en) | 2022-08-30 |
Family
ID=76927593
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110591514.9A Active CN113177367B (en) | 2021-05-28 | 2021-05-28 | High-energy-efficiency federal learning method and device, edge server and user equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113177367B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113406974A (en) * | 2021-08-19 | 2021-09-17 | 南京航空航天大学 | Learning and resource joint optimization method for unmanned aerial vehicle cluster federal learning |
| CN113506040A (en) * | 2021-08-05 | 2021-10-15 | 重庆大学 | Scheduling and resource allocation method for high-performance federal learning in UAV swarm |
| CN114040425A (en) * | 2021-11-17 | 2022-02-11 | 中国电信集团系统集成有限责任公司 | Resource allocation method based on global resource availability optimization |
| CN114118437A (en) * | 2021-09-30 | 2022-03-01 | 电子科技大学 | Model updating synchronization method for distributed machine learning in micro cloud |
| CN114462573A (en) * | 2022-01-20 | 2022-05-10 | 内蒙古工业大学 | Efficient hierarchical parameter transmission delay optimization method oriented to edge intelligence |
| 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 |
| WO2023109828A1 (en) * | 2021-12-15 | 2023-06-22 | 维沃移动通信有限公司 | Data collection method and apparatus, and first device and second device |
| CN117251276A (en) * | 2023-11-20 | 2023-12-19 | 清华大学 | Flexible scheduling method and device for collaborative learning platform |
| CN117635924A (en) * | 2024-01-25 | 2024-03-01 | 南京慧然科技有限公司 | Low-energy-consumption target detection method based on adaptive knowledge distillation |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| US11017322B1 (en) * | 2021-01-28 | 2021-05-25 | Alipay Labs (singapore) Pte. Ltd. | Method and system for federated learning |
-
2021
- 2021-05-28 CN CN202110591514.9A patent/CN113177367B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| US11017322B1 (en) * | 2021-01-28 | 2021-05-25 | Alipay Labs (singapore) Pte. Ltd. | Method and system for federated learning |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113506040A (en) * | 2021-08-05 | 2021-10-15 | 重庆大学 | Scheduling and resource allocation method for high-performance federal learning in UAV swarm |
| CN113506040B (en) * | 2021-08-05 | 2022-06-03 | 重庆大学 | Scheduling and resource allocation method for high-performance federal learning in UAV swarm |
| CN113406974A (en) * | 2021-08-19 | 2021-09-17 | 南京航空航天大学 | Learning and resource joint optimization method for unmanned aerial vehicle cluster federal learning |
| CN113406974B (en) * | 2021-08-19 | 2021-11-02 | 南京航空航天大学 | Learning and resource joint optimization method for unmanned aerial vehicle cluster federal learning |
| CN114118437A (en) * | 2021-09-30 | 2022-03-01 | 电子科技大学 | Model updating synchronization method for distributed machine learning in micro cloud |
| CN114040425A (en) * | 2021-11-17 | 2022-02-11 | 中国电信集团系统集成有限责任公司 | Resource allocation method based on global resource availability optimization |
| CN114040425B (en) * | 2021-11-17 | 2024-03-15 | 中电信数智科技有限公司 | Resource allocation method based on global resource utility rate optimization |
| WO2023109828A1 (en) * | 2021-12-15 | 2023-06-22 | 维沃移动通信有限公司 | Data collection method and apparatus, and first device and second device |
| CN114462573B (en) * | 2022-01-20 | 2023-11-14 | 内蒙古工业大学 | Edge intelligence-oriented efficient hierarchical parameter transmission delay optimization method |
| CN114462573A (en) * | 2022-01-20 | 2022-05-10 | 内蒙古工业大学 | Efficient hierarchical parameter transmission delay optimization method oriented to edge intelligence |
| 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 |
| CN117251276A (en) * | 2023-11-20 | 2023-12-19 | 清华大学 | Flexible scheduling method and device for collaborative learning platform |
| CN117251276B (en) * | 2023-11-20 | 2024-02-09 | 清华大学 | Flexible scheduling method and device for collaborative learning platform |
| CN117635924A (en) * | 2024-01-25 | 2024-03-01 | 南京慧然科技有限公司 | Low-energy-consumption target detection method based on adaptive knowledge distillation |
| CN117635924B (en) * | 2024-01-25 | 2024-05-07 | 南京慧然科技有限公司 | Low-energy-consumption target detection method based on adaptive knowledge distillation |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113177367B (en) | 2022-08-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113177367B (en) | High-energy-efficiency federal learning method and device, edge server and user equipment | |
| CN113516250B (en) | Federal learning method, device, equipment and storage medium | |
| CN113242568B (en) | Task unloading and resource allocation method in uncertain network environment | |
| CN112367109B (en) | Incentive method for digital twin-driven federal learning in air-ground network | |
| CN110928654B (en) | Distributed online task unloading scheduling method in edge computing system | |
| CN112105062B (en) | Mobile edge computing network energy consumption minimization strategy method under time-sensitive condition | |
| CN110557287B (en) | Resource allocation method and device based on Lyapunov optimization | |
| CN113810233B (en) | Distributed computation unloading method based on computation network cooperation in random network | |
| Oren et al. | Online (budgeted) social choice | |
| CN116187483A (en) | Model training method, device, apparatus, medium and program product | |
| CN114169543A (en) | Federal learning algorithm based on model obsolescence and user participation perception | |
| CN116366576A (en) | Method, device, equipment and medium for scheduling computing power network resources | |
| CN116321255A (en) | Compression and user scheduling method for high-timeliness model in wireless federal learning | |
| CN109951540B (en) | Data acquisition method and device based on content timeliness and electronic equipment | |
| Yu et al. | Energy-aware device scheduling for joint federated learning in edge-assisted internet of agriculture things | |
| CN113988160A (en) | Semi-asynchronous layered federal learning updating method based on timeliness | |
| CN116074403B (en) | Multi-protocol fusion communication method, device and architecture | |
| CN104581215A (en) | Content scheduling method and system for internet protocol television (IPTV) | |
| CN109474664B (en) | Active pre-caching method and device in heterogeneous wireless network | |
| CN116486192A (en) | Federal learning method and system based on deep reinforcement learning | |
| Chen et al. | Two-Phase Deep Reinforcement Learning of Dynamic Resource Allocation and Client Selection for Hierarchical Federated Learning | |
| Arvidsson et al. | Optimised local caching in cellular mobile networks | |
| CN110022377B (en) | Adjustment method and system of cloud game server | |
| CN113783798A (en) | Data transmission method and system and edge service equipment | |
| WO2024108601A2 (en) | Terminal selection method, apparatus and system, and model training method, apparatus and system |
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 |