CN116321219B - Self-adaptive honeycomb base station federation forming method, federation learning method and device - Google Patents

Self-adaptive honeycomb base station federation forming method, federation learning method and device Download PDF

Info

Publication number
CN116321219B
CN116321219B CN202310024302.1A CN202310024302A CN116321219B CN 116321219 B CN116321219 B CN 116321219B CN 202310024302 A CN202310024302 A CN 202310024302A CN 116321219 B CN116321219 B CN 116321219B
Authority
CN
China
Prior art keywords
base station
federation
federal
particle
leader
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.)
Active
Application number
CN202310024302.1A
Other languages
Chinese (zh)
Other versions
CN116321219A (en
Inventor
林尚静
李子怡
庄琲
李月颖
张春红
朱新宁
胡铮
陈远祥
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
Original Assignee
Beijing University of Posts and Telecommunications
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 filed Critical Beijing University of Posts and Telecommunications
Priority to CN202310024302.1A priority Critical patent/CN116321219B/en
Publication of CN116321219A publication Critical patent/CN116321219A/en
Application granted granted Critical
Publication of CN116321219B publication Critical patent/CN116321219B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04WWIRELESS COMMUNICATION NETWORKS
    • H04W24/00Supervisory, monitoring or testing arrangements
    • H04W24/02Arrangements for optimising operational condition
    • YGENERAL 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
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02DCLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00Reducing energy consumption in communication networks
    • Y02D30/70Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

The invention provides a self-adaptive cellular base station federation forming method, a federation learning method and a federation learning device, which are characterized in that initial clustering is carried out according to the distribution of base station data to form an initial federation, a federation leader base station is selected in the initial federation by adopting a centrality principle, a federation participant base station is dynamically selected by the federation leader base station through a particle swarm optimization algorithm, and a fitness value is calculated and solved by introducing performance parameters in a federation learning task assignment process, so that the selected federation participant base station can realize the overall optimal federation learning performance, the data which are not independently distributed among cellular base stations in mobile communication can be effectively utilized, and the federation learning task is completed.

Description

Self-adaptive honeycomb base station federation forming method, federation learning method and device
Technical Field
The present invention relates to the field of wireless communications technologies, and in particular, to a federal formation method, a federal learning method, and an apparatus for a self-adaptive cellular base station.
Background
With the high development of mobile communication technology, the mobile communication market represented by smart phones and smart applications has exploded, and the load of the communication network will remain high. In recent years, with the rising and wide application of big data technology, the big data analysis of the mobile communication network is receiving a great deal of attention, and federal learning is gradually applied to the mobile communication network to improve the service quality of the mobile communication and realize more intelligent application.
The existing research applies federal learning to the field of mobile communication, wherein federal learning is a distributed machine learning training framework capable of privacy protection, and a plurality of clients jointly train a data model under the cooperation of a central server. The distributed algorithm architecture proposed by federal learning performs parallelized modeling (channel estimation, traffic prediction, buffer prediction, etc.) on a cooperative wide-range base station, which brings the following advantages: firstly, modeling is performed in cooperation with a large-scale base station, and the prediction precision can be improved due to the increase of data sets; secondly, the training architecture of federal learning parallelization can be maintained, and meanwhile, the difficult problems of high complexity and low real-time performance can not occur.
However, when federal learning is applied to the field of mobile communication, there may be a problem in that the training effect after fusion is poor when model parameters are fused. This is because the conventional federal learning algorithm can obtain a good training effect, which is based on the premise that data samples on a plurality of clients participating in federal learning are independently distributed (IID). However, this premise is not necessarily true in a scenario where it is applied to mobile communication (channel estimation, traffic prediction, buffer prediction, etc.). For example, in a traffic prediction scenario, the data between cellular base stations is in a non-independent co-distribution (non-IID); in a channel estimation scenario, the data of different channels are different; in the cache prediction scene, the requirements of users in different areas on video services are also different, so that the users do not accord with independent same distribution.
In order to solve the problem of data non-IID in federal learning, the existing methods are divided into three categories:
1. And (5) clustering. In the prior art, clustering algorithms such as Gaussian clustering, density-based clustering and K-means clustering are adopted to cluster base stations in the urban global scope, the base stations with similar geographic positions are clustered together for federal training, and geographically adjacent cellular base stations also tend to form federal because of similar data characteristics. However, merely considering geographical locations does not guarantee training results.
2. Personalized federation, existing research proposes unsupervised federation learning based on Dual Averaging (DA) to solve the problem of data non-IID. And calculating the data centroid during clustering by adopting a DA-based method, and calculating the weight of the client. According to the method, a better training effect is obtained, the training accuracy is improved, but higher computational complexity is brought.
3. Data redundancy is introduced. Some existing methods obtain data redundancy by overlapping collection or exchange of raw data sets between clients and introduce data redundancy into the system to handle non-IID data. The method of introducing data redundancy can improve the accuracy of federal learning training, but can increase storage cost and cause the increase of communication cost.
Disclosure of Invention
In view of this, the embodiments of the present invention provide a method for forming federal of self-adaptive cellular base stations, a federal learning method and a federal learning device, so as to eliminate or improve one or more drawbacks existing in the prior art, and solve the problem that the federal learning effect is poor due to the fact that data of each cellular base station is in non-independent distribution in a mobile communication scenario.
The technical scheme of the invention is as follows:
In one aspect, the present invention provides a method for forming a federally adaptive cellular base station, comprising:
Initializing a first cellular base station network topology according to a preset standard according to the space distance of each base station in a set space range;
Introducing divergence to define the difference of data distribution among all base stations, constructing a divergence matrix, and carrying out Hadamard product on the first cellular base station network topology and the divergence matrix to update the first cellular base station network topology to a second cellular base station network topology;
Clustering the network topology of the second cellular base station by adopting a preset clustering algorithm, and clustering the base stations with similar data distribution characteristics to obtain an initial federation cluster of the base stations, wherein the initial federation cluster of the base stations comprises a first set number of initial federations of the base stations;
Each base station initial federation is used for selecting the base station of the federation leader of each base station in a round-by-round manner according to various principles in the centrality principle; in the initial federation of each base station, a federation leader base station of corresponding rounds selects a plurality of federation participant base stations which realize the optimal overall performance of federation learning from the initial federation of the corresponding base stations based on a particle swarm optimization algorithm; each federation leader base station and a corresponding federation participant base station are used for executing federation learning tasks and obtaining performance parameters of the federation learning tasks, wherein the performance parameters are used for calculating fitness values of particles to find historical optimal solutions in the particle swarm optimization algorithm solving process.
In some embodiments, initializing the first cellular base station network topology according to the preset standard according to the spatial distance of each base station in the set spatial range includes:
mapping cellular base station geographical distribution to Wherein v= (V 1,…,vi,vj,…,vN) represents a node corresponding to each base station in the communication network with the set spatial range;
constructing a weight matrix E of edges between corresponding nodes of each base station, wherein the expression is as follows:
where e ij represents the weight of the edge between nodes v i and v j; d ij represents the distance between v i and v j; when d ij is smaller than the threshold value, e ij=1/dij represents that a connecting edge exists between the two nodes; otherwise, if e ij =0, it indicates that there is no edge between the two nodes.
In some embodiments, introducing a divergence to define a difference in data distribution between base stations, constructing a divergence matrix, and performing hadamard product on the first cellular base station network topology and the divergence matrix to update the first cellular base station network topology to a second cellular base station network topology, including:
adopting JS divergence to define the difference of data distribution among all base stations;
for radio traffic data, the hourly radio traffic X traffic of the base station i is a continuous random variable which is quantized and converted into a discrete random variable The probability distribution of wireless traffic per hour for base station i is P i=[pi(x1),…,pi(xs),…,pi(xS), where/>Representing the probability of the radio flow value x s per hour for base station i, and/>The wireless flow Y traffic of the base station j per hour is a continuous random variable, and the continuous random variable is quantized and converted into discrete random variable/> The probability distribution of radio traffic per hour for base station j is P j=[pj(x1),…,pj(xs),…,pj(xS), where/>Represents the probability of a non-current measurement value x s per hour for base station j, an
The difference between the hourly wireless traffic probability distribution P i for base station i and the hourly wireless traffic probability distribution P j for base station j is measured by JS divergence:
for the number of cache contents, the number of cache contents per hour of the base station i is a random variable, the value is X cach={x1,...,xn,...,xI, the probability density distribution of the number of cache contents per hour of the base station i is P i=[pi(x1),…,pi(xn),…,pi(xI), wherein P i(xn)=P(Xcach=xn) represents the probability of the value of the number of cache contents of the base station i of X n, and The number of the cache contents per hour of the base station j is a random variable, the value is Y cach={x1,...,xn,...,xI, the probability distribution of the number of the cache contents per hour of the base station j is P j=[pj(x1),…,pj(xn),…,pj(xI), wherein P j(xn)=P(Ycach=xn) represents the probability of the number of the cache contents per hour of the base station j taking the value of x n, and
The difference between the probability distribution P i of the number of cache contents per hour of the base station i and the probability distribution P j of the number of cache contents per hour of the base station j is measured by JS divergence:
The expression of the divergence matrix is:
The second cellular base station network topology is expressed as:
wherein the symbols are Representing the hadamard product.
In some embodiments, the centrality principle includes at least centrality, mid-number centrality, tight centrality, and feature vector centrality;
The method comprises the steps that a federal leader base station selected based on centrality is a centrality base station, and the centrality base station is connected with the largest number of neighbor cell base stations;
The federal leader base station selected based on the betweenness centrality is the betweenness centrality base station, and the shortest paths from the betweenness centrality base station to other base stations are the most in the connected edges of all the base stations;
The federal leader base station selected based on tight centrality is a tight centrality base station having the shortest path to all other base stations with the shortest average path length;
The federal leader base station selected based on the characteristic vector centrality is a characteristic vector centrality base station, and the characteristic vector centrality base station is connected with the neighbor base station and the secondary neighbor base stations connected with the neighbor base stations, so that the number of the secondary neighbor base stations is the largest.
In some embodiments, in each base station initial federation, selecting, by a federation leader base station of a corresponding round, a plurality of federation participant base stations in a corresponding base station initial federation that achieve optimal overall performance of federation learning based on a particle swarm optimization algorithm, including:
Defining a solution and a solution space for selecting a federal participant base station, wherein for the federal leader base station B m and the set of potential federal participant base stations B cad={b1,…,bm-1,bm+1,…,bM, if there is a federal participant base station/selected, it is marked as x l =1, otherwise, x l =0; the solution of the optimal federal participant base station screening problem is expressed as a vector
Defining a particle p m,k, said particle p m,k having four properties Representing the coordinate position of the particle in solution space,/>Representing the velocity of the particles,/>Representing a historically optimal solution for the particle,/>A fitness value representing the particle; using K particle compositions to search the particle group P m={pm,1,…,pm,K of the optimal federal participant base station screening problem;
Any one particle p m,K∈Pm; in iteration r, the particle p m,k is located at the coordinate position in the solution space Wherein/>A binary variable, which indicates whether the 1 st base station participates or does not participate in the federation in the r-th iteration, the participation is marked as 1, and the non-participation is marked as 0;
in iteration round r, the velocity of particle p m,k The velocity of a particle refers to the position/>, from the r-th wheel coordinate, of the particleCoordinate position/>, when moving to the (r+1) th roundIs a Euclidean distance of (2);
In the r round of iteration, the kth particle The calculation formula of the fitness value of (2) is as follows:
wherein, Representing a benefit for the federal leader base station b m in accordance with the kth particle of the nth roundIndicating the performance parameters obtained after the member base stations participating in the federal training perform the federal training; the performance parameter is a prediction error in a flow prediction scene, is accurate in a channel estimation scene, and is a cache hit rate in a cache prediction scene;
Representing a communication overhead for the federal leader base station b m according to round r, particle k/> The indicated member base stations participating in the federal training carry out the quantity of parameters required to be transmitted after the federal training;
Representing a communication delay for the federal leader base station b m according to the kth particle/>, of the nth round The indicated member base stations participating in the federal training need communication time delay after performing the federal training;
alpha and beta are weight coefficients;
and searching the particles and the solution with the highest particle swarm fitness value according to a particle swarm optimization algorithm.
In some embodiments, the communication overhead is calculated as:
wherein W represents the parameter quantity of the local model of the base station of the federal participant, U represents the parameter fusion times required to be carried out in one federal learning task, Indicating the position of the kth particle of the nth wheel;
The calculation formula of the communication time delay is as follows:
wherein, Representing the position of the kth particle of the nth round,/> Representing the shortest path distance from each base station in the set of potential federal participant base stations B cad to the federal leader base station B m.
In some embodiments, searching for a solution with a highest fitness value for the particle and the particle swarm according to a particle swarm optimization algorithm comprises:
For a pair of Initialization/>And/>
In the r-th round of iteration, all K particles in particle swarm P m update the velocity in parallelThe update formula is:
wherein ω is an inertia constant representing the effect of the speed of the previous iteration particle on the speed of the r-th iteration particle; epsilon 1、ε2 are acceleration constants respectively, and represent the pushing of particles to individual historic optimal solutions respectively Push to population history optimal solution/>Acceleration weights of (2); /(I)Is two random numbers randomly generated in the r-th round,/>
In the r-th iteration process, all K particles in the particle swarm P m update positions in parallel, and the update formula is as follows:
wherein,
In the r-th iteration process, the federation leading base station b m schedules federation member base stations to perform federation training according to the solutions of the current K particle searches, and acquires performance parameters of the federation learning task; calculating the fitness value of the kth particle according to the performance parameters of the federal learning task, and acquiring a historical optimal solution and a population historical optimal solution of the kth particle from the stop to the nth round; the historical optimal solution expression from the kth particle to the nth round is as follows:
the expression of the group history optimal solution is as follows:
After the set round of iterations, a solution of the optimal federal participant base station under the federal leading base station b m is obtained.
In another aspect, the present invention also provides a federal learning method of an adaptive cellular base station, including:
acquiring the space distance and the base station data sequence of each base station in a set space range;
Dynamically screening out a federal leader base station and a plurality of federal participant base stations by adopting the self-adaptive cellular base station federal formation method;
and performing a federal learning task based on the federal leader base station and a plurality of federal participant base stations.
In another aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor implements the steps of the above method when executing the program.
In another aspect, the present invention also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of the above method.
The invention has the advantages that:
According to the self-adaptive cellular base station federation forming method, federation learning method and device, initial clustering is carried out according to the distribution of base station data to form an initial federation, a federation leader base station is selected in the initial federation by adopting a centrality principle, a federation participant base station is dynamically selected by the federation leader base station through a particle swarm optimization algorithm, and a performance parameter calculation fitness value in a federation learning task assignment process is introduced to solve, so that the selected federation participant base station can realize overall optimal federation learning performance, the data which are not independently distributed among cellular base stations in mobile communication can be effectively utilized, and a federation learning task is completed.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings.
It will be appreciated by those skilled in the art that the objects and advantages that can be achieved with the present invention are not limited to the above-described specific ones, and that the above and other objects that can be achieved with the present invention will be more clearly understood from the following detailed description.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate and together with the description serve to explain the application. In the drawings:
fig. 1 is a flow chart of a method for federally forming an adaptive cellular base station according to an embodiment of the present invention.
Fig. 2 is an initial federal schematic diagram of a base station constructed in a federal formation method for an adaptive cellular base station according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the following embodiments and the accompanying drawings, in order to make the objects, technical solutions and advantages of the present invention more apparent. The exemplary embodiments of the present invention and the descriptions thereof are used herein to explain the present invention, but are not intended to limit the invention.
It should be noted here that, in order to avoid obscuring the present invention due to unnecessary details, only structures and/or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details not greatly related to the present invention are omitted.
It should be emphasized that the term "comprises/comprising" when used herein is taken to specify the presence of stated features, elements, steps or components, but does not preclude the presence or addition of one or more other features, elements, steps or components.
It is also noted herein that the term "coupled" may refer to not only a direct connection, but also an indirect connection in which an intermediate is present, unless otherwise specified.
The cellular base station has properties in both the temporal and spatial dimensions. First, spatially, even though base stations are geographically contiguous, there is a discrepancy in data distribution between base stations that cannot be distinguished based on geographical location alone. The base stations with similar geographic positions are gathered together to form the federation, and the problem that the base stations form the federation and the effect is not ideal after participating in federal training is still likely to be faced. Accordingly, in one aspect the present application is directed to solving the problem of base station data samples non-IID. Second, the base station data at adjacent time points is not closely related in time, and the base station data samples are time-varying. Thus, base stations with similar data distribution are gathered together to participate in federal training, and although this method can improve federal learning performance, federal consisting of the previous base stations may collapse due to time-varying characteristics of the base station data, so that the base stations are required to dynamically adjust participants of the federal according to their own data distribution characteristics. Thus, in another aspect the present application addresses the adaptation problems faced by federal formation.
Specifically, the present invention provides a method for forming a federal self-adaptive cellular base station, as shown in fig. 1, including steps S101 to S104:
Step S101: initializing the network topology of the first cellular base station according to a preset standard according to the space distance of each base station in the set space range.
Step S102: introducing divergence to define the difference of data distribution among all base stations, constructing a divergence matrix, and carrying out Hadamard product on the network topology of the first cellular base station and the divergence matrix to update the network topology of the second cellular base station.
Step S103: clustering the network topology of the second cellular base station by adopting a preset clustering algorithm, and clustering the base stations with similar data distribution characteristics to obtain an initial federation cluster of the base stations, wherein the initial federation cluster of the base stations comprises a first set number of initial federations of the base stations.
Step S104: each base station initial federation is used for selecting the base station of the federation leader of each base station in a round-by-round manner according to various principles in the centrality principle; in the initial federation of each base station, a federation leader base station of corresponding rounds selects a plurality of federation participant base stations which realize the optimal overall performance of federation learning from the initial federation of the corresponding base stations based on a particle swarm optimization algorithm; each federation leader base station and the corresponding federation participant base station are used for executing federation learning tasks and obtaining performance parameters of the federation learning tasks, wherein the performance parameters are used for calculating fitness values of particles to find historical optimal solutions in the particle swarm optimization algorithm solving process.
In step S101, the set spatial range may be set according to the requirements of a specific application scenario, and the method may be implemented for a base station in a global scale range of a certain city, or may be implemented by a base station in a spatial range that may be scheduled and utilized continuously. Further, a first cellular base station network topology is initialized to obtain spatial relationships between the base stations.
In some embodiments, initializing the first cellular base station network topology according to the preset standard according to the spatial distance of each base station in the set spatial range includes: mapping cellular base station geographical distribution toWhere v= (V 1,…,vi,vj,…,vN) represents a node corresponding to each base station in the communication network in which the spatial range is set.
Constructing a weight matrix E of edges between corresponding nodes of each base station, wherein the expression is as follows:
Where e ij represents the weight of the edge between nodes v i and v j. d ij represents the distance between v i and v j; when d ij is smaller than the threshold value, e ij=1/dij represents that a connecting edge exists between the two nodes; otherwise, if e ij =0, it indicates that there is no edge between the two nodes.
In step S102, a divergence definition is introduced and differences between the base station data sequences are measured, and the connection closeness between different cellular base station nodes is calculated. KL divergence, also known as relative entropy, information divergence, or information gain, is a measure of asymmetry to the difference in two probability distributions. The invention adopts the variation JS divergence of KL divergence to measure the difference of data distribution among different base stations.
Specifically, in some embodiments, introducing a divergence to define a difference in data distribution between the base stations, constructing a divergence matrix, and performing hadamard product on the first cellular base station network topology and the divergence matrix to update the first cellular base station network topology to a second cellular base station network topology, including:
adopting JS divergence to define the difference of data distribution among all base stations;
for radio traffic data, the hourly radio traffic X traffic of the base station i is a continuous random variable which is quantized and converted into a discrete random variable The probability distribution of wireless traffic per hour for base station i is P i=[pi(x1),…,pi(xs),…,pi(xS), where/>Representing the probability of the radio flow value x s per hour for base station i, and/>The wireless flow Y traffic of the base station j per hour is a continuous random variable, and the continuous random variable is quantized and converted into discrete random variable/> The probability distribution of radio traffic per hour of base station j is P j=[pj(x1),…,pj(xs),…,pj(xS), wherein,Represents the probability of a base station j per hour, a non-current measurement value x s, and/>
The difference between the hourly wireless traffic probability distribution P i for base station i and the hourly wireless traffic probability distribution P j for base station j is measured by JS divergence:
for the number of cache contents, the number of cache contents per hour of the base station i is a random variable, the value is X cach={x1,...,xn,...,xI, the probability density distribution of the number of cache contents per hour of the base station i is P i=[pi(x1),…,pi(xn),…,pi(xI), wherein P i(xn)=P(Xcach=xn) represents the probability of the value of the number of cache contents of the base station i of X n, and The number of the cache contents per hour of the base station j is a random variable, the value is Y cach={x1,...,xn,...,xI, the probability distribution of the number of the cache contents per hour of the base station j is P j=[pj(x1),…,pj(xn),…,pj(xI), wherein P j(xn)=P(Ycach=xn) represents the probability of the number of the cache contents per hour of the base station j taking the value of x n, and
The difference between the probability distribution P i of the number of cache contents per hour of the base station i and the probability distribution P j of the number of cache contents per hour of the base station j is measured by JS divergence:
The expression of the divergence matrix is:
The second cellular base station network topology is expressed as:
wherein the symbols are Representing the hadamard product.
In step S103, in the second cellular base station network topology, toThe method represents the weight among the base stations, the edge weight value among the base stations with closer distances and more similar data distribution is higher, and the edge weight value among the base stations with farther distances and more different data distribution is lower. Based on the weight, the edge weight sum between sub-graphs corresponding to the initial federal cluster of the base station obtained by dividing can be made as low as possible, and the edge weight sum in the sub-graphs is as high as possible. Specifically, the clustering can adopt k-means, k-modes and other algorithms.
In step S104, for each initial federation cluster of base stations screened in step S103, the base stations in the cluster have similar data distribution characteristics, in order to make federation learning obtain better effects, in this embodiment, further select a federation leader base station in each initial federation cluster of base stations, and screen a plurality of federation participant base stations in the corresponding initial federation cluster of base stations by the federation leader base station, so as to make the overall performance of federation learning optimal.
The process of screening the base stations of the federation participants by the base stations of the federation leader mainly adopts a particle swarm optimization algorithm, and in the process of solving by utilizing the particle swarm optimization algorithm, the process can be divided into an inner loop and an outer loop, wherein the outer loop adopts different centrality principles to select the base stations of the federation leader, and the inner loop carries out screening of the base stations of the federation participants under the condition that the base stations of the federation leader are established.
In some embodiments, the centrality principle includes at least centrality, mid-order centrality, tight centrality, and feature vector centrality.
The federal leader base station selected based on the centrality is the centrality base station, and the centrality base station is connected with the largest number of neighbor cell base stations.
The federal leader base station selected based on the medium center is the medium center base station, and the shortest paths from the medium center base station to other base stations are the largest in the connected edges of all the base stations.
The federal leader base station selected based on tight centrality is the tight centrality base station, which has the shortest path length to all other base stations on average.
The federal leader base station selected based on the feature vector centrality is the feature vector centrality base station, and the number of neighbor base stations connected by the feature vector centrality base station and secondary neighbor base stations connected by each neighbor base station is the largest.
Of course, the centrality principle is not limited to the four above, it being understood that other types of centrality principles may be used to select a federal leader base station. In other embodiments, the federal leader base station may be pre-designated, or a trusted or power-compliant base station may be selected as the federal leader base station according to the needs of the application. After determining the federal leader base station in one round of the outer loop, the federal participant base stations are screened in the inner loop using a particle optimization algorithm. In the process of screening the particle optimization algorithm by using the particle optimization algorithm, the particle corresponding base station of each round is utilized to calculate the fitness of the overall performance obtained after federal learning so as to guide the updating of the particles and the particle swarm.
In some embodiments, in each initial federation of base stations, selecting, by a federation leader base station of a corresponding round, a plurality of federation participant base stations that achieve the best overall performance of federation learning from the initial federation of the corresponding base stations based on a particle swarm optimization algorithm, including steps S201 to S20:
Step S201: defining a solution and a solution space for selecting a federal participant base station, wherein for a federal leader base station B m and a set of potential federal participant base stations B cad={b1,…,bm-1,bm+1,…,bM, if there is a federal participant base station l selected, it is marked as x l =1, otherwise, x l =0; the solution of the optimal federal participant base station screening problem is expressed as a vector
Step S202: defining particle p m,k, particle p m,k has four properties Representing the coordinate position of the particle in the solution space,/>Representing the velocity of the particles,/>Represents the historical optimal solution of the particle,A fitness value representing the particle; the K particle swarm P m={pm,1,…,pm,K used to search for the optimal federal participant base station screening problem is composed.
Step S203: any one particle p m,K∈Pm; in iteration r, the particle p m,k is located at the coordinate position in the solution spaceWherein/>Is a binary variable indicating whether the 1 st base station participates or does not participate in the federation in the r-th iteration, the participation being noted as 1 and the non-participation being noted as 0.
In iteration round r, the velocity of particle p m,k The velocity of a particle refers to the position/>, from the r-th wheel coordinate, of the particleCoordinate position/>, when moving to the (r+1) th roundIs a euclidean distance of (c).
In the r round of iteration, the kth particleThe calculation formula of the fitness value of (2) is as follows:
wherein, alpha and beta are weight coefficients. Representing the benefit of the federal leader base station b m according to the kth particle/>, of the nth roundIndicating the performance parameters obtained after the member base stations participating in the federal training perform the federal training; the performance parameter is a prediction error in a traffic prediction scene, accuracy in a channel estimation scene, and cache hit rate in a cache prediction scene.
Representing communication overhead for the federal leader base station b m according to the kth particle of the nth roundThe indicated member base stations participating in the federal training have the number of parameters to be transmitted after the federal training. In some embodiments, the communication overhead is calculated as:
Wherein W represents the parameter quantity of the local model of the base station of the federal participant, U represents the parameter fusion times required to be carried out in one federal learning task, Indicating the position of the kth particle of the nth wheel;
Representing the communication time delay, wherein the communication time delay is that the base station b m of the federal leader is based on the kth particle/>, of the nth round And the indicated member base stations participating in the federal training need communication delay after performing the federal training. In some embodiments, the communication latency is calculated as:
wherein, Representing the position of the kth particle of the nth round,/> Representing the shortest path distance from each base station in the set of potential federal participant base stations B cad to the federal leader base station B m.
Step S204: and searching a solution with the highest fitness value of the particles and the particle swarm according to the particle swarm optimization algorithm.
In some embodiments, searching for a solution with highest fitness value between particles and particle swarm according to the particle swarm optimization algorithm includes steps S301-S303:
step S301: for a pair of Initialization/>And/>
Step S302: in the r-th round of iteration, all K particles in particle swarm P m update the velocity in parallelThe update formula is:
wherein ω is an inertia constant representing the effect of the speed of the previous iteration particle on the speed of the r-th iteration particle; epsilon 1、ε2 are acceleration constants respectively, and represent the pushing of particles to individual historic optimal solutions respectively Push to population history optimal solution/>Acceleration weights of (2); /(I)Is the two random numbers randomly generated in the r-th round, and aims to increase the randomness of search,/>
Step S303: in the r-th iteration process, all K particles in the particle swarm P m update positions in parallel, and the update formula is as follows:
wherein,
Step S304: in the r round of iteration process, the federation leading base station b m schedules federation member base stations to perform federation training according to the solutions of the current K particle searches, and acquires performance parameters of federation learning tasks; calculating the fitness value of the kth particle according to the performance parameters of the federal learning task, and obtaining the historical optimal solution and the population historical optimal solution of the kth particle from the stop to the nth round; the historical optimal solution expression from the kth particle to the nth round is as follows:
the expression of the population history optimal solution is:
Step S304: after the set round of iterations, a solution of the optimal federal participant base station under the federal leading base station b m is obtained.
On the other hand, the invention also provides a federal learning method of the self-adaptive cellular base station, which comprises the following steps of S401 to S403:
Step S401: and acquiring the space distance and the base station data sequence of each base station in the set space range.
Step S402: the adaptive cellular base station federation formation method is adopted to dynamically screen out a federal leader base station and a plurality of federal participant base stations.
Step S403: the federal learning task is performed based on the federal leader base station and a plurality of federal participant base stations.
In another aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor implements the steps of the above method when executing the program.
In another aspect, the present invention also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of the above method.
The invention is described below in connection with a specific embodiment:
The embodiment provides a federal formation method of a self-adaptive cellular base station. The method comprises three steps: firstly, according to the base stations in the urban global scale range, a cellular base station network topology is constructed according to the distances between the base stations and the data sample distribution of the base stations, geographic positions and data distribution are considered, and clustering algorithm is adopted to cluster the base station topology to form an initial base station federal cluster. Next, for each initial base station federation, a base station is promoted as a federal leader according to different criteria (centrality criteria), such as centrality criteria for degree centrality, median centrality, etc. The method aims at ensuring a shorter average inter-node communication distance between the leader base station and the participant base station, so as to ensure that the communication overhead between the leader base station and the participant base station is small and the transmission delay is short. The federal leader is responsible for parameter fusion during federal training and screens participant base stations in the federation under its leader. And finally, the federal leader base station screens the participant base stations in the federation under the leader to form the optimal base station federation. Specifically, the federal leader base station dynamically searches for the optimal federal participant base station through a particle swarm optimization algorithm (PARTICLE SWARM Optimizing, PSO) so that the performance of all members of the federal base station under the self-leader after federal learning is optimal.
A first part: initial cellular base station federal formation
Step 1: user plane functions (User Plane Function, UPF) network elements in the core network build the initial federation of base stations. Aiming at the base stations in the urban global scale range, the user plane network elements form a cellular base station network topology according to the data distribution similarity of the base stations and the distance between the base stations, and clustering the base station topology by using a clustering algorithm to construct an initial federal cluster of the base stations. The specific steps are as follows:
step 1.1: a complex network topology of cellular base stations is built. The user plane network element regards the cellular base stations as network nodes, the connection edge between the nodes refers to the connection relation (optical fiber connection exists) between two cellular base stations, and the geographical distribution of the cellular base stations is mapped into a graph Wherein v= (V 1,…,vi,vj,…,vN) represents all nodes in the network, N is the total number of nodes in the graph, and a weight matrix E of edges between corresponding nodes of each base station is constructed to represent all connected edges in the network, and the following formula is shown:
Where e ij represents the weight of the edge between nodes v i and v j. d ij represents the distance between v i and v j. When d ij is smaller than the threshold value, e ij=1/dij represents that a connecting edge exists between the two nodes; otherwise, if e ij =0, it indicates that there is no edge between the two nodes. The number of neighbor cell sites that have a border with a cell site is taken as the site node degree.
Step 1.2: on the basis of defining the network topology of the cellular base station in step 11, the user plane network element introduces divergence to recalculate the degree of closeness of the connection between the nodes of the cellular base station, forming a new topology structure of the cellular base station. Divergence is introduced to define the variability of data distribution between cell base station i and neighbor cell base station j. The JS divergence is defined as d JS(Pi||Pj),Pi and P j, which represent the data distribution of base station i and base station j, respectively. The smaller the JS divergence, the higher the similarity of the two distributions.
For radio traffic data, the hourly radio traffic X traffic of the base station i is a continuous random variable which is quantized and converted into a discrete random variableThe probability distribution of wireless traffic per hour for base station i is P i=[pi(x1),…,pi(xs),…,pi(xS), where/>Representing the probability of the radio flow value x s per hour for base station i, and/>The wireless flow Y traffic of the base station j per hour is a continuous random variable, and the continuous random variable is quantized and converted into discrete random variable/> The probability distribution of radio traffic per hour of base station j is P j=[pj(x1),…,pj(xs),…,pj(xS), wherein,Represents the probability of a base station j per hour, a non-current measurement value x s, and/>
The difference between the hourly wireless traffic probability distribution P i for base station i and the hourly wireless traffic probability distribution P j for base station j is measured by JS divergence:
for the number of cache contents, the number of cache contents per hour of the base station i is a random variable, the value is X cach={x1,...,xn,...,xI, the probability density distribution of the number of cache contents per hour of the base station i is P i=[pi(x1),…,pi(xn),…,pi(xI), wherein P i(xn)=P(Xcach=xn) represents the probability of the value of the number of cache contents of the base station i of X n, and The number of the cache contents per hour of the base station j is a random variable, the value is Y cach={x1,...,xn,...,xI, the probability distribution of the number of the cache contents per hour of the base station j is P j=[pj(x1),…,pj(xn),…,pj(xI), wherein P j(xn)=P(Ycach=xn) represents the probability of the number of the cache contents per hour of the base station j taking the value of x n, and
The difference between the probability distribution P i of the number of cache contents per hour of the base station i and the probability distribution P j of the number of cache contents per hour of the base station j is measured by JS divergence:
The expression of the divergence matrix is:
multiplying the cellular base station topology constructed in the step 1.1 by the same positions of the JS divergence matrix respectively to form a new cellular base station network topology on the basis of considering the distance between the cellular base stations and the similarity of data distribution Wherein, symbol/>Representing the hadamard product.
Step 1.3: based on the new cellular base station network topology constructed in step 1.2, the user plane network element utilizes a clustering algorithm to construct the base station federation, so as to form an initial cellular base station federation, as shown in fig. 2. Representing the weights of the edges between the base stations asThe base stations with closer distances and more similar data distribution have higher edge weight values, and the base stations with farther distances and more different data distribution have lower edge weight values. And cutting the updated cellular base station topological graph to ensure that the edge weight between the obtained subgraphs is as low as possible and the edge weight in the subgraphs is as high as possible, thereby achieving the purpose of clustering the base stations with similar data distribution characteristics together.
A second part: adaptive cellular base station federal formation
The base stations with similar data distribution characteristics are clustered together in the first part to form N initial federal clusters f= { F 1,…,Fn,…,FN }. The present embodiment is described taking the nth initial base station federation F n in the initial base station federation cluster F as an example. Suppose that in total, in initial base station federal F n, it is composed of M geographically adjacent cellular base stations, i.e. Taking the initial federation F n of the base station as an example, the formation process of the self-adaptive base station federation is described, and for convenience of description, in the subsequent description process, the base station members/>, in the base station federation are ignoredThe superscript n of (b m) indicates the mth base station in the nth initial base station association.
Step 2.1: and (5) promoting the federal leader base station of the federal of the cellular base station.
Base station initial federal F n proposes a base station as a federal leader base station according to different criteria, and the federal leader base station screens federal participant base stations in its federal under its leader. The base station initial federal F n promotes the federal leader on a central basis. The centrality criteria include: center of degree, center of betweenness, center of tightness, center of feature vector. The reason for adopting the centrality criterion is that firstly, the central base station in the initial federation of the base station has a pivot position in a geographic space, global information is easier to acquire, and the base station naturally has a federal leader position; second, there is typically a short average inter-node communication distance between the central base station in the initial federal of base stations and the federal participant base stations. Therefore, the communication overhead between the nodes is small, and the transmission delay is short.
Step 2.1.1: the base station initial federal F n sequentially promotes the federal leader base station by the centrality criteria such as degree centrality, medium centrality, tight centrality, feature vector centrality, etc. Assuming T centrality criteria, a total of T rounds of base station federal leader election is performed.
A isocentric base station refers to the base station itself having the greatest degree (number of neighbor cellular base stations).
The medium center base station refers to the base station with the largest shortest path number from the base station to other base stations in the continuous edge of all base stations.
A tight centrality base station refers to the base station having the shortest path length from the base station to all other base stations.
The feature vector centrality base station refers to that the base station has larger degree (the number of the neighbor cellular base stations) per se, and the degree of the neighbor nodes is larger than that of other base stations.
Step 2.1.2: in the t (t.gtoreq.1) th round of federal leader base station election, it is assumed that base station initial federal F n elects mth base station b m as the federal leader base station.
Step 2.1.3: during the t-th round of federal leader base station election, the federal leader base station b m is responsible for screening the federal participant base stations under its leader. The federal leader base station b m decides whether it adds or does not add to the base station federation under its own leader for all M-1 base stations except itself for the federal participant set { b 1,…,bm-1,bm+1,…,bM } in the base station's initial federation, which is a 0-1 knapsack problem. Thus, the federal leader base station b m can search for the optimal federal participant base station by a particle swarm optimization algorithm (PARTICLE SWARM Optimizing, PSO) so that the overall performance of all members of the base station federation under its own leader after federal learning is optimal. The process of searching for the optimal federal participant base station is detailed in step 2.2.
Step 2.1.4: and iterating t=t+1 until t=t, and obtaining the leader base station after T rounds of iteration and the optimal federal participant base station obtained by screening under the corresponding leader base station.
Step 2.1.5: and taking the base station federation corresponding to the solution with the maximum fitness value as the final optimal base station federation according to the fitness value of the solution of the T base station federations obtained by screening, wherein the definition of the fitness value is detailed in the step 2.3.
Step 2.2: and forming the federal of the cellular base stations.
Step 2.2.1: and defining a solution and a solution space of the optimal federal member screening problem. The federal leader base station B m randomly picks a federal participant base station from the potential federal participant base station set B cad={b1,…,bm-1,bm+1,…,bM. If base station b l,bl∈Bcad is selected to participate in federation, x l =1, otherwise x l =0. Then a solution to the optimal federal participant base station screening problem can be represented by a vector, i.eThe solution space refers to the space of all feasible solutions, namely, one vector space formed by the set of all feasible solutions of the optimal federal participant screening problem.
Step 2.2.2: particles and particle clusters are defined that search for the optimal federal member screening problem described above. Defining the particle belonging to the federally leadership b m as p m,k, the particle p m,k having four attribute values,The federally-leaded base station b m uses K particle compositions to search the particle population P m={pm,1,…,pm,K for the optimal federal member screening problem described above.
Step 2.3: the federal leader base station b m initiates a total of one particle population P m of K particles in parallel with R-round iterations to search for the optimal federal participant base station combination.
In the r-th round of iteration, the following iteration is performed for any one particle p m,K∈Pm.
Coordinate position of the r-th round iterative particle p m,k in solution spaceWherein/>Is a binary variable indicating whether the 1 st base station participates or does not participate in the federation in the r-th iteration.
Speed of the iterative particle p m,k of the r-th roundThe velocity of a particle refers to the position/>, from the r-th wheel coordinate, of the particleCoordinate position/>, when moving to the (r+1) th roundIs a distance of (3). The distance is defined herein as the Euclidean distance.
The (th) round of iteration (kth) particleThe fitness function of (a) is defined as follows:
wherein, First item/>Representing benefits, second term/>Representing communication overhead, third item/>And the communication delay is represented, and alpha and beta are weight coefficients. /(I)Indicating that federal leader base station b m is according to round r, kth particle/>After the indicated member base stations participating in the federal training perform the federal training, the overall performance is obtained, and the specific federal training process is shown in the third section. Representing a prediction error in a traffic prediction scene; in the channel estimation scene, representing accuracy; in the context of cache prediction, the cache hit rate is indicated. /(I)Can be defined as:
representing federal leader base station b m according to round r, kth particle/> After the indicated member base stations participating in the federal training perform the federal training, the required communication overhead can be expressed as: /(I)
Wherein W represents the parameter quantity of the local model of the base station of the federal participant, U represents the parameter fusion times required to be carried out in one federal learning task,Expressed as/>Indicating the position of the kth particle of the nth wheel.
Representing federal leader base station b m according to round r, kth particle/>After the indicated member base stations participating in the federal training perform the federal training, the required communication delay can be expressed as:
wherein, Expressed as/>Representing the position of the kth particle of the nth round,/>Representing the shortest path distance from each base station in the set of potential federal participant base stations B cad to the federal leader base station B m.
By the r-th round, the expression of the historic optimal solution searched by the particle p m,k is as follows:
And stopping the process until the r-th round, and searching the solution with the highest fitness value by the particle p m,k.
By the r-th round, the historical optimal solution expression searched by the particle swarm P m is as follows:
And stopping the process until the r-th round, and searching the solution with the highest fitness value by all K particles.
Step 2.4: the specific search algorithm is as follows.
Step 2.4.1: for a pair ofInitialization/>
Step 2.4.2: in the iterative process of the (r.gtoreq.1) th round, all K particles in the particle group P m update the speed in parallelAt iteration round r, the velocity of particle p m,k/>Is mainly updated by the r-1 round speed/>Individual experience of particle p m,k to round r, and population experience of particle p m,k to round r.
Wherein the particle p m,k is accumulated to the individual experience of the r round, which means the solution searched by the particle p m,k at the r roundWith self-searched historical optimal solution/>Distance between/>Group experience of particle p m,k accumulated to round r, which means solution searched by round r of particle p m,k/>Historical optimal solution searched with particle swarm P m/>Distance between/>Specifically, the following formula is shown:
Where ω is the inertia constant and represents the effect of the velocity of the previous iteration particle on the velocity of the iteration particle of the r-th round. Epsilon 1、ε2 are acceleration constants respectively, and represent the pushing of particles to individual historic optimal solutions respectively Push to population history optimal solution/>Is a weight of acceleration of the system. /(I)Is the two random numbers randomly generated in the r-th round, and aims to increase the randomness of search,/>
Step 2.4.3: in the iterative process of the (r.gtoreq.1) th round, all K particles in the particle swarm P m update positions in parallel, namely
Wherein,
Due to the above mentionedIs a binary variable, the position of the particle before updating/>Is an integer variable, and therefore a switching function needs to be introduced to convert the position of the particles into a variable consisting of 0-1 in order to determine whether the base station is involved in federal training. The switching function can be defined as when the particle position/>And when the output is smaller than the threshold value, the output is 0, otherwise, the output is 1.
Step 2.4.4: in the process of the r (r is more than or equal to 1) round of iteration, the federal leading base station b m schedules federal member base stations to perform federal training according to the solution of the current K particle searches.
(1) The solution currently searched for the kth particle p k If/>Then the member base station b l participates in federal training, otherwise, the member base station b l does not participate in federal training. In round r, particle/>The indicated member base stations participate in federal training, and the leader base station b m is responsible for model fusion in the federal learning process, while also performing local model training.
(2) The larger the fitness function value is, the kth particle of the nth wheel is representedThe better the performance. By comparing the magnitude of the fitness value, updating the historical optimal solution of the kth particle until the kth round is finished, and the expression is as follows:
(3) k=k+1 until the maximum number of iterations K is reached. Comparing the K particle history optimal solutions, and updating the group history optimal solution of the K particles until the number of the r is reached
Step 2.4.5: iteration, r=r+1, until r=r, yields the optimal base station federation under the leader base station B m after R iterations, denoted B atd={bl∈Bcad|xl =1.
Third section: federal training process for base station participants
Step 3.1: assume that the federal participant base station model training needs to perform U-wheel federal parameter fusion, and E epochs are trained in each wheel, and U is trained in a total. And after the maximum fusion times are reached, all the federal participant base stations stop updating the parameters of the model.
Step 3.2: u=0, a global model parameter W (0) is initialized at the federal leader base station node and pushed to the individual federal participant base stations. Referring to the description of the second section, federal participant base station set B atd under leader base station B m, there are a total of |b atd |federal participant base stations. All federal participant base stations begin the federal training process based on the unified initial model parameters.
Step 3.2.1: on the u-th round, the federal participant base station b l,bl∈Batd receives the global model parameters Wu -1, let B l training according to the local data to obtain the local model parameters/>, after the u th round of updating, of each round of updating parameters of the federal participant base station
(A) When epoch=0, the training sample set for federal participant base station b l is expressed as: there are a total of S training batches, i.e./> Initial s=1, calculate b l's loss function on each training batch, the loss function of the s-th batch is/>
(B) For each training batch, the obtained loss function is utilized to update the local model parameters according to the gradient descent method,Wherein/>Representing the loss function/>, of federal participant base station b l over the s batches at the time of training round tEta represents the learning rate, eta e 0, 1.
(C) s=s+1, letRepeating steps (a) and (b) until s=s, i.e. all batches of samples are trained. At this time, 1 epoch training is completed at base station participant b l to obtain local model parameters
(D) epoch=epoch+1, letRepeating steps (a) to (c) until epoch=e. Base station participant b l obtains the final model parameters of the current round of updates/>
Step 3.2.2: all |B atd | federal participant base stations upload the trained model parameters to the leader base station node, the leader base station node performs primary parameter fusion on the global model according to a federal average algorithm or a weighted federal average algorithm, the fused global model parameters are expressed as Wu,Where Wu represents the global model after the u-th round of updating.
Step 3.2.3: the leader base station node pushes the global model parameters Wu after the u-th round of fusion to |B atd | base station participants.
Step 3.3: u=u+1, repeating steps 3.2.1 to 3.2.3 until the maximum number of fusions U is reached.
According to the method for adaptively forming the cell base station federation, the base stations are initially clustered by combining the distribution of the base station data samples, and the leader base station is enabled to dynamically screen the participant base stations by adopting a PSO algorithm, so that the optimal base station federation is formed. The adaptive cellular base station federal formation method improves the accuracy of federal learning training of non-IID, and has originality in the field of mobile communication.
In summary, according to the self-adaptive cellular base station federation forming method, federation learning method and device, initial clustering is performed according to the distribution of base station data to form an initial federation, a federation leader base station is selected in the initial federation by adopting a centrality principle, a federation participant base station is dynamically selected by the federation leader base station through a particle swarm optimization algorithm, and a performance parameter calculation fitness value in a federation learning task assignment process is introduced to solve, so that the selected federation participant base station can realize overall optimization of federation learning performance, and data in non-independent distribution among cellular base stations in mobile communication can be effectively utilized to complete federation learning tasks.
Those of ordinary skill in the art will appreciate that the various illustrative components, systems, and methods described in connection with the embodiments disclosed herein can be implemented as hardware, software, or a combination of both. The particular implementation is hardware or software dependent on the specific application of the solution and the design constraints. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention. When implemented in hardware, it may be, for example, an electronic circuit, an Application Specific Integrated Circuit (ASIC), suitable firmware, a plug-in, a function card, or the like. When implemented in software, the elements of the invention are the programs or code segments used to perform the required tasks. The program or code segments may be stored in a machine readable medium or transmitted over transmission media or communication links by a data signal carried in a carrier wave. A "machine-readable medium" may include any medium that can store or transfer information. Examples of machine-readable media include electronic circuitry, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio Frequency (RF) links, and the like. The code segments may be downloaded via computer networks such as the internet, intranets, etc.
It should also be noted that the exemplary embodiments mentioned in this disclosure describe some methods or systems based on a series of steps or devices. The present invention is not limited to the order of the above-described steps, that is, the steps may be performed in the order mentioned in the embodiments, or may be performed in a different order from the order in the embodiments, or several steps may be performed simultaneously.
In this disclosure, features that are described and/or illustrated with respect to one embodiment may be used in the same way or in a similar way in one or more other embodiments and/or in combination with or instead of the features of the other embodiments.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, and various modifications and variations can be made to the embodiments of the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (8)

1. An adaptive cellular base station federal formation method, comprising:
Initializing a first cellular base station network topology according to a preset standard according to the space distance of each base station in a set space range;
Introducing divergence to define the difference of data distribution among all base stations, constructing a divergence matrix, and carrying out Hadamard product on the first cellular base station network topology and the divergence matrix to update the first cellular base station network topology to a second cellular base station network topology;
Clustering the network topology of the second cellular base station by adopting a preset clustering algorithm, and clustering the base stations with similar data distribution characteristics to obtain an initial federation cluster of the base stations, wherein the initial federation cluster of the base stations comprises a first set number of initial federations of the base stations;
each base station initial federation is used for selecting a federation leader base station of each base station initial federation in a round-by-round manner according to various principles in a centrality principle; in the initial federation of each base station, selecting a feasible federation participant base station by using a federation leader base station of a corresponding round based on the current state of particles in a particle swarm algorithm to execute a federation learning task, and acquiring performance parameters of the federation learning task, wherein the performance parameters are used for calculating the fitness of the particles to update the state of the particles so as to search the optimal federation participant base station of the federation leader base station of the corresponding round; in the initial federation of each base station, comparing the overall performances of each round of federation leader base station and the optimal federation participant base station, and determining the optimal federation leader base station and the optimal federation participant base station;
The method comprises the steps of initializing a first cellular base station network topology according to a preset standard and the space distance of each base station in a set space range, wherein the method comprises the following steps:
mapping cellular base station geographical distribution to Wherein v= (V 1,…,vi,vj,…,vN) represents a node corresponding to each base station in the communication network with the set spatial range;
constructing a weight matrix E of edges between corresponding nodes of each base station, wherein the expression is as follows:
Where e ij represents the weight of the edge between nodes v i and v j; d ij represents the distance between v i and v j; when d ij is smaller than the threshold value, e ij=1/dij represents that a connecting edge exists between the two nodes; otherwise, if e ij =0, it indicates that there is no edge between the two nodes;
The centrality principle at least comprises centrality, median centrality, tight centrality and characteristic vector centrality; the method comprises the steps that a federal leader base station selected based on centrality is a centrality base station, and the centrality base station is connected with the largest number of neighbor cell base stations; the federal leader base station selected based on the betweenness centrality is the betweenness centrality base station, and the shortest paths from the betweenness centrality base station to other base stations are the most in the connected edges of all the base stations; the federal leader base station selected based on tight centrality is a tight centrality base station having the shortest path to all other base stations with the shortest average path length; the federal leader base station selected based on the characteristic vector centrality is a characteristic vector centrality base station, and the characteristic vector centrality base station is connected with the neighbor base station and the secondary neighbor base stations connected with the neighbor base stations, so that the number of the secondary neighbor base stations is the largest.
2. The adaptive cellular base station federation formation method of claim 1, wherein introducing a divergence defining a difference in data distribution between base stations, constructing a divergence matrix, and performing hadamard product on the first cellular base station network topology and the divergence matrix to update to a second cellular base station network topology, comprises:
adopting JS divergence to define the difference of data distribution among all base stations;
for radio traffic data, the hourly radio traffic X traffic of the base station i is a continuous random variable which is quantized and converted into a discrete random variable The probability distribution of wireless traffic per hour for base station i is P i=[pi(x1),…,pi(xs),…,pi(xS), where/> Representing the probability of the radio flow value x s per hour for base station i, and/>The wireless flow Y traffic of the base station j per hour is a continuous random variable, and the continuous random variable is quantized and converted into discrete random variable/>The probability distribution of radio traffic per hour of base station j is P j=[pj(x1),…,pj(xs),…,pj(xS), wherein,Represents the probability of a base station j per hour, a non-current measurement value x s, and/>
The difference between the hourly wireless traffic probability distribution P i for base station i and the hourly wireless traffic probability distribution P j for base station j is measured by JS divergence:
for the number of cache contents, the number of cache contents per hour of the base station i is a random variable, the value is X cach={x1,...,xn,...,xI, the probability density distribution of the number of cache contents per hour of the base station i is P i=[pi(x1),…,pi(xn),…,pi(xI), wherein P i(xn)=P(Xcach=xn) represents the probability of the value of the number of cache contents of the base station i of X n, and The number of the cache contents per hour of the base station j is a random variable, the value is Y cach={x1,...,xn,...,xI, the probability distribution of the number of the cache contents per hour of the base station j is P j=[pj(x1),…,pj(xn),…,pj(xI), wherein P j(xn)=P(Ycac=xn) represents the probability of the number of the cache contents per hour of the base station j taking the value of x n, and
The difference between the probability distribution P i of the number of cache contents per hour of the base station i and the probability distribution P j of the number of cache contents per hour of the base station j is measured by JS divergence:
The expression of the divergence matrix is:
The second cellular base station network topology is expressed as:
wherein the symbols are Representing the hadamard product.
3. The adaptive cellular base station federation formation method according to claim 1, wherein in each base station initial federation, a federation leader base station of a corresponding round selects a plurality of federation participant base stations achieving optimum federation learning overall performance from the corresponding base station initial federation based on a particle swarm optimization algorithm, comprising:
Defining a solution and a solution space for selecting a federal participant base station, wherein for the federal leader base station B m and the set of potential federal participant base stations B cad={b1,…,bm-1,bm+1,…,bM, if there is a federal participant base station/selected, it is marked as x l =1, otherwise, x l =0; the solution of the optimal federal participant base station screening problem is expressed as a vector
Defining a particle p m,k, said particle p m,k having four propertiesRepresenting the coordinate position of the particle in solution space,/>Representing the velocity of the particles,/>Representing a historical optimal solution for the particle,A fitness value representing the particle; using K particle compositions to search the particle group P m={pm,1,…,pm,K of the optimal federal participant base station screening problem;
Any one particle p m,K∈Pm; in iteration r, the particle p m,k is located at the coordinate position in the solution space Wherein/>A binary variable, which indicates whether the 1 st base station participates or does not participate in the federation in the r-th iteration, the participation is marked as 1, and the non-participation is marked as 0;
in iteration round r, the velocity of particle p m,k The velocity of a particle refers to the position/>, from the r-th wheel coordinate, of the particleCoordinate position/>, when moving to the (r+1) th roundIs a Euclidean distance of (2);
In the r round of iteration, the kth particle The calculation formula of the fitness value of (2) is as follows:
wherein, Representing a benefit for the federal leader base station b m according to round r, kth particle/>Indicating the performance parameters obtained after the member base stations participating in the federal training perform the federal training; the performance parameter is a prediction error in a flow prediction scene, is accurate in a channel estimation scene, and is a cache hit rate in a cache prediction scene;
Representing communication overhead for the federal leader base station b m according to round r, particle k/> The indicated member base stations participating in the federal training carry out the quantity of parameters required to be transmitted after the federal training;
Representing a communication delay for the federal leader base station b m according to the kth particle of the nth round The indicated member base stations participating in the federal training need communication time delay after performing the federal training;
alpha and beta are weight coefficients;
and searching the particles and the solution with the highest particle swarm fitness value according to a particle swarm optimization algorithm.
4. The adaptive cellular base station federation formation method according to claim 3, wherein the communication overhead calculation formula is:
wherein W represents the parameter quantity of the local model of the base station of the federal participant, U represents the parameter fusion times required to be carried out in one federal learning task, Indicating the position of the kth particle of the nth wheel;
the calculation formula of the communication time delay is as follows:
wherein, Indicating the position of the kth particle of the nth wheel,Representing the shortest path distance from each base station in the set of potential federal participant base stations B cad to the federal leader base station B m.
5. The adaptive cellular base station federation formation method according to claim 4, wherein searching for the highest fitness value solution of the particles and the particle swarm according to a particle swarm optimization algorithm comprises:
For a pair of Initialization/>And/>
In the r-th round of iteration, all K particles in particle swarm P m update the velocity in parallelThe update formula is:
wherein ω is an inertia constant representing the effect of the speed of the previous iteration particle on the speed of the r-th iteration particle; epsilon 1、ε2 are acceleration constants respectively, and represent the pushing of particles to individual historic optimal solutions respectively Pushing to population history optimal solutionsAcceleration weights of (2); /(I)Is two random numbers randomly generated in the r-th round,/>
In the r-th iteration process, all K particles in the particle swarm P m update positions in parallel, and the update formula is as follows:
wherein,
In the r-th iteration process, the federation leading base station b m schedules federation member base stations to perform federation training according to the solutions of the current K particle searches, and acquires performance parameters of the federation learning task; calculating the fitness value of the kth particle according to the performance parameters of the federal learning task, and acquiring a historical optimal solution and a population historical optimal solution of the kth particle from the stop to the nth round; the historical optimal solution expression from the kth particle to the nth round is as follows:
the expression of the group history optimal solution is as follows:
After the set round of iterations, a solution of the optimal federal participant base station under the federal leading base station b m is obtained.
6. A method of federal learning an adaptive cellular base station, comprising:
acquiring the space distance and the base station data sequence of each base station in a set space range;
Dynamically screening out a federal leader base station and a plurality of federal participant base stations using the adaptive cellular base station federal formation method of any one of claims 1 to 5;
and performing a federal learning task based on the federal leader base station and a plurality of federal participant base stations.
7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 6 when the program is executed.
8. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 6.
CN202310024302.1A 2023-01-09 2023-01-09 Self-adaptive honeycomb base station federation forming method, federation learning method and device Active CN116321219B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310024302.1A CN116321219B (en) 2023-01-09 2023-01-09 Self-adaptive honeycomb base station federation forming method, federation learning method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310024302.1A CN116321219B (en) 2023-01-09 2023-01-09 Self-adaptive honeycomb base station federation forming method, federation learning method and device

Publications (2)

Publication Number Publication Date
CN116321219A CN116321219A (en) 2023-06-23
CN116321219B true CN116321219B (en) 2024-04-19

Family

ID=86796761

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310024302.1A Active CN116321219B (en) 2023-01-09 2023-01-09 Self-adaptive honeycomb base station federation forming method, federation learning method and device

Country Status (1)

Country Link
CN (1) CN116321219B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113660112A (en) * 2021-07-23 2021-11-16 广州大学 Bandwidth allocation method, system, computer equipment and medium for federated learning
CN114091617A (en) * 2021-11-29 2022-02-25 深圳前海微众银行股份有限公司 Federal learning modeling optimization method, electronic device, storage medium, and program product
CN114143891A (en) * 2021-11-30 2022-03-04 南京工业大学 FDQL-based multi-dimensional resource collaborative optimization method in mobile edge network
WO2022051964A1 (en) * 2020-09-10 2022-03-17 Qualcomm Incorporated Reporting for information aggregation in federated learning
CN115580891A (en) * 2022-12-09 2023-01-06 北京邮电大学 Flow prediction model training method, prediction method and device based on federal learning

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11907403B2 (en) * 2021-06-10 2024-02-20 Hong Kong Applied Science And Technology Research Institute Co., Ltd. Dynamic differential privacy to federated learning systems

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022051964A1 (en) * 2020-09-10 2022-03-17 Qualcomm Incorporated Reporting for information aggregation in federated learning
CN113660112A (en) * 2021-07-23 2021-11-16 广州大学 Bandwidth allocation method, system, computer equipment and medium for federated learning
CN114091617A (en) * 2021-11-29 2022-02-25 深圳前海微众银行股份有限公司 Federal learning modeling optimization method, electronic device, storage medium, and program product
CN114143891A (en) * 2021-11-30 2022-03-04 南京工业大学 FDQL-based multi-dimensional resource collaborative optimization method in mobile edge network
CN115580891A (en) * 2022-12-09 2023-01-06 北京邮电大学 Flow prediction model training method, prediction method and device based on federal learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Optimizing Multi-Objective Federated Learning on Non-IID Data with Improved NSGA-III and Hierarchical Clustering;Jialin Zhong 等;MDPI Symmetry 2022;20220523;第1-19页 *
Reducing Communication Costs of Federated Contrastive Learning by Particle Swarm Optimization;Zesheng Liu 等;IAECST 2021;20220207;第687-692页 *
基于个性化联邦学习的无线通信流量预测;林尚静 等;电 子 科 技 大 学 学 报;20230118;第52卷(第1期);第64-73页 *

Also Published As

Publication number Publication date
CN116321219A (en) 2023-06-23

Similar Documents

Publication Publication Date Title
CN112902969B (en) Path planning method of unmanned aerial vehicle in data collection process
CN110968426A (en) Edge cloud collaborative k-means clustering model optimization method based on online learning
CN109831236B (en) Beam selection method based on Monte Carlo tree search assistance
CN108737569B (en) Service selection method facing mobile edge computing environment
CN108668287B (en) Active caching method based on user content popularity and mobile rule
Chua et al. Resource allocation for mobile metaverse with the Internet of Vehicles over 6G wireless communications: A deep reinforcement learning approach
CN114629930A (en) Method and communication device for semi-asynchronous federal learning
CN115065678A (en) Multi-intelligent-device task unloading decision method based on deep reinforcement learning
CN110809893A (en) Electronic device and method for wireless communication
CN109673015B (en) Tracking area planning method based on spectral clustering
CN111757266A (en) UAV data acquisition trajectory algorithm based on solar power supply type agricultural Internet of things
CN108810139B (en) Monte Carlo tree search-assisted wireless caching method
Sheena et al. Multi‐objective metaheuristic optimization‐based clustering with network slicing technique for Internet of Things‐enabled wireless sensor networks in 5G systems
CN114268348A (en) Honeycomb-free large-scale MIMO power distribution method based on deep reinforcement learning
Yi et al. DMADRL: A distributed multi-agent deep reinforcement learning algorithm for cognitive offloading in dynamic MEC networks
Han et al. Dynamic task offloading and service migration optimization in edge networks
CN116321219B (en) Self-adaptive honeycomb base station federation forming method, federation learning method and device
CN116484976A (en) Asynchronous federal learning method in wireless network
Gad et al. Communication-efficient federated learning in drone-assisted IoT networks: Path planning and enhanced knowledge distillation techniques
Yan et al. Regression-based K nearest neighbours for resource allocation in network slicing
CN116634450A (en) Dynamic air-ground heterogeneous network user association enhancement method based on reinforcement learning
Si et al. UAV-assisted Semantic Communication with Hybrid Action Reinforcement Learning
Ozturk et al. Introducing a novel minimum accuracy concept for predictive mobility management schemes
Nie et al. Joint access point fuzzy rough set reduction and multisource information fusion for indoor Wi-Fi positioning
CN104955059B (en) Cellular network base stations state time-varying model method for building up based on Bayesian network

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