CN113791895A - Edge calculation and resource optimization method based on federal learning - Google Patents
Edge calculation and resource optimization method based on federal learning Download PDFInfo
- Publication number
- CN113791895A CN113791895A CN202110958468.1A CN202110958468A CN113791895A CN 113791895 A CN113791895 A CN 113791895A CN 202110958468 A CN202110958468 A CN 202110958468A CN 113791895 A CN113791895 A CN 113791895A
- Authority
- CN
- China
- Prior art keywords
- user
- edge
- model
- users
- training
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000005457 optimization Methods 0.000 title claims abstract description 25
- 238000004364 calculation method Methods 0.000 title claims abstract description 20
- 230000006870 function Effects 0.000 claims abstract description 53
- 238000012549 training Methods 0.000 claims abstract description 45
- 238000004891 communication Methods 0.000 claims abstract description 30
- 238000005265 energy consumption Methods 0.000 claims abstract description 25
- 230000008569 process Effects 0.000 claims abstract description 25
- 230000005540 biological transmission Effects 0.000 claims description 47
- 230000002776 aggregation Effects 0.000 claims description 37
- 238000004220 aggregation Methods 0.000 claims description 37
- 238000013468 resource allocation Methods 0.000 claims description 10
- 238000006116 polymerization reaction Methods 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 7
- 238000012935 Averaging Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000012417 linear regression Methods 0.000 claims description 4
- 238000012821 model calculation Methods 0.000 claims description 4
- 238000003491 array Methods 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 5
- 230000007246 mechanism Effects 0.000 abstract description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 3
- 230000003190 augmentative effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5072—Grid computing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- 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
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computing Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses an edge calculation and resource optimization method based on federal learning, which is characterized in that an antenna array is deployed at a small base station, channel information of a downlink and channel information processed by a precoding technology are obtained, channel and precoding are formed as training data of an input-output pair, and federal learning is carried out under the support of the data, namely, a model is trained at a node end, and finally, the purpose of inputting the channel information to obtain corresponding precoding information is achieved. In the process, in order to obtain a stable learning alliance and control the system energy consumption in the lowest state, user selection is carried out, namely, a plurality of users participate in training by selecting users with stable computing capability and communication capability through the physical characteristics of each node; in order to encourage users to actively participate in training, a contract mechanism is introduced to reward the users participating in training, the income and the paid training cost of each user are calculated to obtain a utility function, and the resources are distributed to the users so that the utility of the whole system is maximized.
Description
Technical Field
The invention belongs to the field of communication and federal learning, combines the precoding technology in MIMO with machine learning, is applied to edge calculation, and enables the federal learning structure to be more stable and efficient by means of selection and resource optimization.
Background
The large-scale Multiple Input Multiple Output (MIMO) technology is to deploy a large-scale array at the end of a base station, and compared with the traditional MIMO, the MIMO technology can effectively resist the interference between different users, and can significantly improve the capacity of the system. In a large-scale MIMO system, a precoding technique is a crucial signal processing technique in a downlink, and converts a modulated symbol stream into a data stream adapted to a current channel by using Channel State Information (CSI) of a transmitting end, so as to concentrate signal energy near a target user, effectively combat attenuation and loss, and improve system performance. The method specifically comprises the steps that a sending end of a downlink utilizes CSI to preprocess a sending signal, interference among different users and antennas is minimized, signal energy is concentrated to the vicinity of a target user, a receiving end obtains a better signal-to-noise ratio (SNR), and the system channel capacity is improved [1 ]. Under the background of machine learning, traditional federal learning is frequently used, receipts of an edge node end are sent to a server end to be processed, then the server end distributes data to a node end to be trained, privacy of the data is greatly reduced, centralized federal learning is proposed and well applied to the intelligent field in order to protect the data, the node does not need to share the data and only needs to be trained locally, in order to improve feasibility of the data, a weight needs to be trained by the data and sent to the server end, the server end plays a role in collection, data weights trained by all nodes are collected and correspondingly processed, then the processed weights are broadcasted to all nodes, and after the nodes receive the latest data weights, the weights are updated and applied to the next iteration, this loops until the data is trained to the best model. And the traditional system architecture is composed of a mobile node side and a base station side [2 ]. The limited computing and traffic capabilities of mobile devices have made machine learning suffer from two major bottlenecks, namely, the powerful functions of communication overhead and energy overhead Mobile Edge Computing (MEC), which is considered a promising distributed computing paradigm that supports many of the 5G era emerging intelligent applications such as video streaming, smart cities, and augmented reality. According to this concept of allowing delay-sensitive and compute-intensive tasks to be offloaded from distributed mobile devices to close-range edge servers, providing real-time response and high energy efficiency, a hierarchical federation edge learning framework is introduced, where edge servers are typically deployed fixed with base stations, as intermediaries between mobile devices and the cloud, can execute local models transmitted from nearby devices for edge aggregation. And each local device achieves the given learning precision, and global aggregation is carried out on edge transmission so as to realize model updating. A reduction in communication and energy costs is achieved in this manner [3], thereby entering a three-tier communication system architecture.
[1] Zhang Yu, Zhao Xiongwen, precoding technique in millimeter wave large scale MIMO system [ J ]. Zhongxing communication technique, 2018,024(003):26-31
[2]Elbir A M,Coleri S.Federated Learning for Hybrid Beamforming in mm-Wave Massive MIMO[J].IEEE Communications Letters,2020,PP(99):1-1.
[3]S.Luo,X.Chen,Q.Wu,Z.Zhou and S.Yu,"HFEL:Joint EdgeAssociation and ResourceAllocation for Cost-Efficient Hierarchical Federated Edge Learning,"in IEEE Transactions on Wireless Communications,vol.19,no.10,pp.6535-6548,Oct.2020,doi:10.1109/TWC.2020.3003744.
Disclosure of Invention
The invention aims to introduce the MEC into a communication system structure for federal learning to form a three-layer structure, namely a macro base station (cloud server), a small base station (edge server) and a node user, and allocate communication resources to the user under the application background of federal learning so that the whole system achieves the maximum effect, and meanwhile, the MIMO precoding technology is linked with the federal learning training data so as to predict corresponding precoding under the condition that channel data is known. Firstly, deploying an antenna array at a small base station, acquiring channel information of a downlink and channel information processed by a precoding technology, forming a channel and precoding as training data of an input-output pair, and performing federal learning under the data support, namely training a model at a node end, and finally achieving the purpose that corresponding precoding information can be obtained by inputting the channel information. In the process, in order to obtain a stable learning alliance and control the system energy consumption in the lowest state, user selection is carried out, namely, a plurality of users participate in training by selecting users with stable computing capability and communication capability through the physical characteristics of each node; meanwhile, in order to encourage users to actively participate in training, a contract mechanism is introduced to reward the users participating in training, the income and the paid training cost of each user are calculated to obtain a utility function (the utility is income-cost), and the resources are allocated to the users so that the utility of the whole system is maximized.
In order to solve the problems, the invention adopts the following technical scheme: the edge calculation and resource optimization method based on the federal learning comprises the following steps:
step 1: establishing a system model and deploying arrays according to users, edge servers and base stations
The invention relates to a multi-user millimeter wave MIMO three-layer structure system structure, wherein the first layer is a user layer and consists of N single-antenna user devices. The second layer is the ES layer (edge server layer) containing K layers with STAn edge server for the antenna. The third layer is a CS layer (cloud server layer), and has a cloud server S with a single antenna. Wherein, the user layer and the ES layer communicate through a wireless channel, and the ES layer and the CS layer are connected through a backhaul link, having a sufficient capacity.
Step 2: a set of users communicating with the edge server is first selected according to distance and represented as
Wherein N is the number of users,
the number of users selected after the distance selection is performed is shown.
That is, assuming that the user equipment obeys a poisson distribution, the communication range of each edge server is a radius dkAnd in this three-level structure, no overlap between the circular communication areas of the edge servers is assumed. And then selecting the users with the shorter distance according to the distance from each user to the edge server, wherein the communication quality between the users and the server in the communication range is better, and stable communication is a necessary condition for ensuring the successful federal learning.
And step 3: and acquiring training data.
The channel model is an LOS (line of sight transmission) receiving route of an L path, and the millimeter wave channel matrix of the nth user is expressed as follows:
wherein the content of the first and second substances,
αn,lis the complex channel gain of the l-th path,
for the directional angle of the downlink transmission path,
is STX 1, whose mth element is represented as:
λ is the wavelength, σ is the distance between the two antennas, and let σ equal λ/2.
Firstly, firstly
Within the range of which the transmission direction angle of the subscriber is generated
Randomly generating L-path channel transmission direction angle in a fluctuation range of (5 degrees), and then solving for h according to a channel matrix formulanAngle interval of
Divided equally into non-overlapping Q parts, each subinterval
And marks each section range starting from 1, a section where the user transmission direction angle falls is marked as q, and the q value is a section mark value, thereby generating (h)nAnd q) data pairs. And the final radio frequency precoding fRF,nCan be expressed as
Wherein
Is the midpoint of the interval.
And 4, step 4: training is performed according to the selected users and the generated MIMO data.
The training process can be divided into edge aggregation and cloud aggregation
1) Edge polymerization: edge clustering is divided into local model computation, local model transmission and edge model aggregation. The specific contents are as follows:
local model calculation: after locally generating MIMO data, user n uses the local data to perform linear regression model training based on gradient descent until local model accuracy theta is reachedn∈[0,1]I.e. the number of iterations is L (theta)n)＝μlog(1/θn) Then, model parameters w are generatedtAnd calculating the loss function f of the modeli(wt). In the process, the calculation time delay is generated
And calculating energy
The computation time delay and the computation energy consumed can be derived from the amount of data trained by the user and the frequency with which the data is processed.
Local model transmission: after the local computation is over, the user n will model the parameters wtUploading to an edge server k in communication with the edge server k, and adopting an OFDMA (orthogonal frequency division multiple access) protocol in the transmission process, wherein the transmission time delay is generated
And transferring energy
The transmission delay and the transmission energy can be calculated by a shannon formula according to the transmission rate between the user n and the edge server k and the bandwidth allocated to the user n by the server k.
Edge polymerization: in the process, the edge server aggregates the model parameters uploaded by each user and then performs an averaging process to update the model parameters (i.e. a federal averaging algorithm). After the edge server carries out model parameter aggregation and an average algorithm, updated model parameters are sent to each user in a broadcasting mode (time delay and energy consumption in the broadcasting process are ignored), after the user receives the model parameters sent by the server, the user uses the parameters to carry out a new round of data training, and the operation is repeated in a circulating mode until the edge server reaches a model precision value epsilon, and the times of global aggregation are expressed as
Where δ is a learning task parameter. When I (epsilon, theta) is performedn) The total delay/energy consumption generated after the sub-global aggregation is the calculated delay/energy consumption and the transmission delayTotal of energy consumption.
2) Cloud polymerization: the cloud aggregation comprises two steps of edge model uploading and cloud model aggregation
Uploading the edge model: after the edge server k reaches the edge precision, the edge model parameters are uploaded to the cloud server, and time delay is generated in the process
And energy consumption
May be calculated from the transmission rate and the transmission power.
Cloud model aggregation: the cloud server aggregation process is similar to the edge aggregation process, namely the cloud server receives the model parameters uploaded by the edge server and then carries out average calculation.
The key points of the invention are as follows:
step 1: allocating resources for users
Secondary selection: in order to further obtain a stable and efficient learning alliance, a second user selection is carried out before resources are allocated to the users, the selection is carried out according to the training record of the user at the previous time, namely before the user selection is carried out, the user needs to carry out local iteration and achieve corresponding precision, then the gradient value is worked out according to the loss function of local calculation, the difference value is carried out with the gradient gathered at the edge server end, and the users with small gradient difference values are selected to be constructed into the final learning alliance
Wherein
Is the set of users after the first selection,
is the set of users after two selections.
Resource allocation: after the second selection, the alliance is successfully established and then participates in the final participationThe resource allocation is carried out by the trained users, in the process, the contract theorem is introduced to encourage and reward the users participating in the training, and the users are firstly classified according to the local precision (the smaller the better), namely
Then sorting is carried out, and simultaneously, an evaluation function of a user is set
Wherein
Is the reward for participating in the league learning user,
is a function of profit and is a function of
Increasing and increasing monotonically increasing functions. Thus, the computational utility function of the user can be obtained, i.e.
Edge server providing contract terms
The user selects proper terms according to the actual situation of the user, and the proper terms are respectively IR (personal rational constraint) and IC (incentive compatibility constraint) according to the constraint conditions of the contract theorem.
IR: the income and cost loss of the user are at the lowest level, i.e.
If u'n<0, the user cannot participate in the training and is discarded.
IC: the user must match himself when selecting contract terms, i.e.
According to the utility function u of the user can be obtainednUtility function U with edge serverk(the difference between the cost function and the income function is ignored, the utility function of the cloud server is ignored, in conclusion, the utility function of the whole system can be obtained
Where K is a set of edge servers,
are the set of users after two selections.
The purpose of resource allocation is to minimize the system communication overhead and maximize the utility function of the system, so an optimization problem can be presented, namely
max U,
subject to,
C1:
C2:
C3:
C4:
C5:
C6:
Wherein the constraint condition C1And C2Respectively representing uplink communication resource constraints and computation capability constraints, betak,nIs the bandwidth allocated to user n by edge server k, fnIs the calculated frequency of user n and
and
is the minimum and maximum of the calculated frequency. C4And C5Ensure the stability and high efficiency of the equipment participating in the model training6It is ensured that the users participating in the training are users within the service range of the edge server.
However, the optimization problem is due to the constraint C4And C5It becomes difficult to solve and the optimization problem is simplified below with respect to these two constraints.
1. For any valid clause
If and only if
When the temperature of the water is higher than the set temperature,
2. for any valid clause
If and only if
When the temperature of the water is higher than the set temperature,
3. for any valid clause
The computational utility function of each type of user must be satisfied
0≤u'1<…<u'm<…<u'n
4. Based on 1 and 2, constraint C5(IC) can be solved as Local Down IC (LDIC), i.e.
And Local Upward IC (LUIC), i.e.
Constraint C according to 1, 2, 3, and 44And C5Can be solved as:
then the constraints of the optimization problem are reduced to:
C1:
C2:
suppose that
And ∈ ≠ 1, from constraint condition C1Can be solved out
From constraint C2Can be solved out
To sum up, can obtain
This is a nonlinear programming optimization problem with constraints, and f can be solved under the condition that the utility function is satisfied to the maximumnAnd betak,nI.e., the optimal value, thereby completing the resource allocation.
The traditional federal learning is generally a two-layer structure, namely a user layer and a server layer, but the distance between the user layer and the server layer is usually longer, so that the transmission delay and the transmission energy consumption are serious, and the communication overhead is greatly increased.
In the invention, in order to strengthen the stability of the training alliance, a user selection link is added, and a user with strong calculation and communication capabilities is selected to be added into the training, so that the calculation time delay and the calculation energy consumption are greatly reduced, and the communication overhead is further reduced.
Drawings
Fig. 1 shows the loss function of the linear regression model calculated after the local precision is reached when a user trains the local MIMO channel data pair using the gradient descent algorithm, and it can be clearly seen from the figure that when the number of iterations reaches 400-500, the loss function converges to a fixed minimum value, which illustrates the situation that the user trains the model to be optimal at this time.
Fig. 2 and 3 show the resource allocation result after the second user selection, and it can be seen from fig. 3 that when the step value is 2, the inverse value F of the utility function of the system reaches the minimum value (i.e. the utility function reaches the maximum value), and at this time, the frequency resource and the bandwidth resource take the optimal values, which can be clearly seen from fig. 2.
Fig. 4 and 5 show the delay and energy consumption of the traditional federal learning (U-M) with a two-tier structure and the federal learning (U-S-M) with a three-tier structure after the edge server is introduced, and under the condition that the number of users and the precision of a user model are the same, the number of users is 15-20, and it can be seen that the delay and energy consumption of the three-tier structure are obviously less than those of the two-tier structure.
Fig. 6 shows the system delay and energy consumption with no user selection and resource allocation, and it can be seen from the figure that under the same condition, the system delay and energy consumption after selection are much smaller than those before selection, the system delay and energy consumption after optimization are smaller than those before optimization, and the system after selection and optimization is in the best state, i.e. the case where the delay and energy consumption are minimum.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and examples.
The edge calculation and resource optimization method based on the federal learning comprises the following steps:
step 1: establishing a system model and deploying arrays according to users, edge servers and base stations
The invention relates to a multi-user millimeter wave MIMO three-layer structure system structure, wherein the first layer is a user layer and consists of N single-antenna user devices. The second layer is the ES layer (edge server layer) containing K layers with STAn edge server for the antenna. The third layer is a CS layer (cloud server layer), and has a cloud server S with a single antenna. Wherein, the user layer and the ES layer communicate through a wireless channel, and the ES layer and the CS layer are connected through a backhaul link, having a sufficient capacity.
Step 2: a set of users communicating with the edge server is first selected according to distance and represented as
Wherein N is the number of users,
the number of users selected after the distance selection is performed is shown.
That is, assuming that the user equipment obeys a poisson distribution, the communication range of each edge server is a radius dkAnd in this three-level structure, no overlap between the circular communication areas of the edge servers is assumed. And then selecting the users with the shorter distance according to the distance from each user to the edge server, wherein the communication quality between the users and the server in the communication range is better, and stable communication is a necessary condition for ensuring the successful federal learning.
And step 3: and acquiring training data.
The channel model is an LOS (line of sight transmission) receiving route of an L path, and the millimeter wave channel matrix of the nth user is expressed as follows:
wherein the content of the first and second substances,
αn,lis the complex channel gain of the l-th path,
for the directional angle of the downlink transmission path,
is STX 1, whose mth element is represented as:
λ is the wavelength, σ is the distance between the two antennas, and let σ equal λ/2.
Firstly, firstly
Within the range of which the transmission direction angle of the subscriber is generated
Randomly generating L-path channel transmission direction angle in a fluctuation range of (5 degrees), and then solving for h according to a channel matrix formulanAngle interval of
Divided equally into non-overlapping Q parts, each subinterval
And marks each section range starting from 1, a section where the user transmission direction angle falls is marked as q, and the q value is a section mark value, thereby generating (h)nAnd q) data pairs. And the final radio frequency precoding fRF,nCan be expressed as
Wherein
Is the midpoint of the interval.
And 4, step 4: training is performed according to the selected users and the generated MIMO data.
The training process can be divided into edge aggregation and cloud aggregation
2) Edge polymerization: edge clustering is divided into local model computation, local model transmission and edge model aggregation. The specific contents are as follows:
local model calculation: after locally generating MIMO data, user n uses the local data to perform linear regression model training based on gradient descent until local model accuracy theta is reachedn∈[0,1]I.e. iterationThe number of times is L (theta)n)＝μlog(1/θn) Then, model parameters w are generatedtAnd calculating the loss function f of the modeli(wt). In the process, the calculation time delay is generated
And calculating energy
The computation time delay and the computation energy consumed can be derived from the amount of data trained by the user and the frequency with which the data is processed.
Local model transmission: after the local computation is over, the user n will model the parameters wtUploading to an edge server k in communication with the edge server k, and adopting an OFDMA (orthogonal frequency division multiple access) protocol in the transmission process, wherein the transmission time delay is generated
And transferring energy
The transmission delay and the transmission energy can be calculated by a shannon formula according to the transmission rate between the user n and the edge server k and the bandwidth allocated to the user n by the server k.
Edge polymerization: in the process, the edge server aggregates the model parameters uploaded by each user and then performs an averaging process to update the model parameters (i.e. a federal averaging algorithm). After the edge server carries out model parameter aggregation and an average algorithm, updated model parameters are sent to each user in a broadcasting mode (time delay and energy consumption in the broadcasting process are ignored), after the user receives the model parameters sent by the server, the user uses the parameters to carry out a new round of data training, and the operation is repeated in a circulating mode until the edge server reaches a model precision value epsilon, and the times of global aggregation are expressed as
Where δ is a learning task parameter. When in useCarrying out I (epsilon, theta)n) The total latency/energy consumption resulting after the sub-global aggregation is the sum of the computational latency/energy consumption and the transmission latency/energy consumption.
2) Cloud polymerization: the cloud aggregation comprises two steps of edge model uploading and cloud model aggregation
Uploading the edge model: after the edge server k reaches the edge precision, the edge model parameters are uploaded to the cloud server, and time delay is generated in the process
And energy consumption
May be calculated from the transmission rate and the transmission power.
Cloud model aggregation: the cloud server aggregation process is similar to the edge aggregation process, namely the cloud server receives the model parameters uploaded by the edge server and then carries out average calculation.
The key points of the invention are as follows:
step 1: allocating resources for users
Secondary selection: in order to further obtain a stable and efficient learning alliance, a second user selection is carried out before resources are allocated to the users, the selection is carried out according to the training record of the user at the previous time, namely before the user selection is carried out, the user needs to carry out local iteration and achieve corresponding precision, then the gradient value is worked out according to the loss function of local calculation, the difference value is carried out with the gradient gathered at the edge server end, and the users with small gradient difference values are selected to be constructed into the final learning alliance
Wherein
Is the set of users after the first selection,
is the set of users after two selections.
Resource allocation: after the second selection, the alliance is successfully established, then the resources of the users who finally participate in the training are distributed, a contract theorem is introduced in the process to encourage and reward the users who participate in the training, and the users are firstly classified according to the local precision (the smaller the better), namely the users are classified according to the local precision (the smaller the better the users are), namely
Then sorting is carried out, and simultaneously, an evaluation function of a user is set
Wherein
Is the reward for participating in the league learning user,
is a function of profit and is a function of
Increasing and increasing monotonically increasing functions. Thus, the computational utility function of the user can be obtained, i.e.
Edge server providing contract terms
The user selects proper terms according to the actual situation of the user, and the proper terms are respectively IR (personal rational constraint) and IC (incentive compatibility constraint) according to the constraint conditions of the contract theorem.
IR: the income and cost loss of the user are at the lowest level, i.e.
If u'n<0, the user cannot participate in the training and is abandoned。
IC: the user must match himself when selecting contract terms, i.e.
According to the utility function u of the user can be obtainednUtility function U with edge serverk(the difference between the cost function and the income function is ignored, the utility function of the cloud server is ignored, in conclusion, the utility function of the whole system can be obtained
Where K is a set of edge servers,
are the set of users after two selections.
The purpose of resource allocation is to minimize the system communication overhead and maximize the utility function of the system, so an optimization problem can be presented, namely
max U,
subject to,
C1:
C2:
C3:
C4:
C5:
C6:
Wherein the constraint condition C1And C2Respectively representing uplink communication resource constraints and computation capability constraints, betak,nIs the bandwidth allocated to user n by edge server k, fnIs the calculated frequency of user n and
and
is the minimum and maximum of the calculated frequency. C4And C5Ensure the stability and high efficiency of the equipment participating in the model training6It is ensured that the users participating in the training are users within the service range of the edge server.
However, the optimization problem is due to the constraint C4And C5It becomes difficult to solve and the optimization problem is simplified below with respect to these two constraints.
3. For any valid clause
If and only if
When the temperature of the water is higher than the set temperature,
4. for any valid clause
If and only if
When the temperature of the water is higher than the set temperature,
3. for any valid clause
The computational utility function of each type of user must be satisfied
0≤u'1<…<u'm<…<u'n
4. Based on 1 and 2, constraint C5(IC) can be solved as Local Down IC (LDIC), i.e.
And Local Upward IC (LUIC), i.e.
Constraint C according to 1, 2, 3, and 44And C5Can be solved as:
then the constraints of the optimization problem are reduced to:
C1:
C2:
suppose that
And ∈ ≠ 1, from constraint condition C1Can be solved out
From constraint C2Can be solved out
To sum up, can obtain
This is a nonlinear programming optimization problem with constraints, and f can be solved under the condition that the utility function is satisfied to the maximumnAnd betak,nI.e., the optimal value, thereby completing the resource allocation.
Claims (5)
1. The edge calculation and resource optimization method based on the federal learning is characterized in that: the method comprises the following steps:
step 1: establishing a system model and deploying arrays according to users, edge servers and base stations
The multi-user millimeter wave MIMO three-layer structure system structure comprises a first layer, a second layer and a third layer, wherein the first layer is a user layer and consists of N single-antenna user devices; the second layer is an ES layer, i.e., an edge server layer, containing K layers with STAn edge server for the antenna; the third layer is a CS layer, namely a cloud server layer, and is provided with a cloud server S with a single antenna; the user layer and the ES layer communicate through a wireless channel, and the ES layer and the CS layer are connected through a backhaul link and have sufficient capacity;
step 2: a set of users communicating with the edge server is first selected according to distance and represented as
Wherein N is the number of users,
the number of the selected users after the distance selection is performed during the representation; that is, assuming that the user equipment obeys a poisson distribution, the communication range of each edge server is a radius dkAnd in this three-level structure, no overlap between the circular communication areas of the edge servers is assumed; then, selecting the users with the closer distance according to the distance from each user to the edge server;
and step 3: acquiring training data;
the channel model is an LOS receiving route of an L path, and the millimeter wave channel matrix of the nth user is expressed as:
wherein the content of the first and second substances,
is the complex channel gain of the l-th path,
for the directional angle of the downlink transmission path,
is STX 1, whose mth element is represented as:
λ is the wavelength, σ is the distance between the two antennas, and let σ equal λ/2;
firstly, firstly
Within the range of which the transmission direction angle of the subscriber is generated
Randomly generating L-path channel transmission direction angle in the fluctuation range, and then solving h according to a channel matrix formulanAngle interval of
Divided equally into non-overlapping Q parts, each subinterval
And marks each section range starting from 1, a section where the user transmission direction angle falls is marked as q, and the q value is a section mark value, thereby generating (h)nQ) data pairs; final radio frequency precoding fRF，nIs shown as
Wherein
Is the midpoint of the interval;
and 4, step 4: training according to the selected user and the generated MIMO data; the training process is divided into edge aggregation and cloud aggregation.
2. The federated learning-based edge computing and resource optimization method of claim 1, wherein: the edge polymerization: the edge aggregation is divided into local model calculation, local model transmission and edge model aggregation; the specific contents are as follows:
local model calculation: after locally generating MIMO data, user n uses the local data to perform linear regression model training based on gradient descent to achieve local model accuracyθn∈[0，1]I.e. the number of iterations is L (theta)n)＝μlog(1/θn) Then, model parameters w are generatedtAnd calculating the loss function f of the modeli(wt) (ii) a In the process, the calculation time delay is generated
And calculating energy
The calculation time delay and the consumed calculation energy can be obtained according to the data volume of the user training and the frequency of processing the data;
local model transmission: after the local computation is over, the user n will model the parameters wtUploading to an edge server k in communication with the edge server k, and adopting an OFDMA (orthogonal frequency division multiple access) protocol in the transmission process, wherein the transmission time delay is generated
And transferring energy
The transmission delay and the transmission energy can be calculated by a Shannon formula according to the transmission rate between the user n and the edge server k and the bandwidth allocated to the user n by the server k;
edge polymerization: in the process, the edge server gathers the model parameters uploaded by each user and then carries out averaging processing so as to update the model parameters (namely, a federal averaging algorithm); after the edge server carries out model parameter aggregation and an average algorithm, updated model parameters are sent to each user in a broadcasting mode (time delay and energy consumption in the broadcasting process are ignored), after the user receives the model parameters sent by the server, the user uses the parameters to carry out a new round of data training, and the operation is repeated in a circulating mode until the edge server reaches a model precision value epsilon, and the times of global aggregation are expressed as
Where δ is a learning task parameter; when I (epsilon, theta) is performedn) The total latency/energy consumption resulting after the sub-global aggregation is the sum of the computational latency/energy consumption and the transmission latency/energy consumption.
3. The federated learning-based edge computing and resource optimization method of claim 1, wherein: the cloud aggregation comprises the following steps: the cloud aggregation comprises two steps of edge model uploading and cloud model aggregation;
uploading the edge model: after the edge server k reaches the edge precision, the edge model parameters are uploaded to the cloud server, and time delay is generated in the process
And energy consumption
Can be calculated according to the transmission rate and the transmission power;
cloud model aggregation: the cloud server aggregation process is similar to the edge aggregation process, namely the cloud server receives the model parameters uploaded by the edge server and then carries out average calculation.
4. The federated learning-based edge computing and resource optimization method of claim 1, wherein: before the user selects, the user needs to carry out local iteration to reach corresponding precision, then the gradient value is worked out according to the loss function calculated locally, and difference value is made between the gradient value and the gradient gathered by the edge server end, and the user with small gradient difference value is selected to be constructed as the final learning alliance
Wherein
Is the set of users after the first selection,
is the set of users after two selections.
5. The federated learning-based edge computing and resource optimization method of claim 1, wherein: after the second selection, the alliance is successfully established, then the resources of the users who finally participate in the training are distributed, a contract theorem is introduced in the process to encourage and reward the users who participate in the training, and the users are firstly classified according to local precision, namely the users are classified according to the local precision
Then sorting is carried out, and simultaneously, an evaluation function of a user is set
Wherein
Is the reward for participating in the league learning user,
is a function of profit and is a function of
A monotonically increasing function that increases; the computational utility function of the user is thus obtained, i.e.
Edge server providing contract terms
The user selects proper terms according to the actual situation of the user, and the terms are respectively IR and IC according to the constraint conditions of the contract theorem;
IR: the income and cost loss of the user are at the lowest level, i.e.
If u'nIf < 0, the user can not participate in the training and is abandoned;
IC: the user must match himself when selecting contract terms, i.e.
Obtaining the utility function u of the user according to the abovenUtility function U with edge serverk(the difference value between the cost function and the income function is ignored by the utility function of the cloud server; in conclusion, the utility function of the whole system is obtained
Where K is a set of edge servers,
is a user set after two selections;
the purpose of resource allocation is to minimize the system communication overhead and maximize the utility function of the system, so an optimization problem is presented, namely
max U，
subject to，
C1：
C2：
C3：
C4：
C5：
C6：
Wherein the constraint condition C1And C2Respectively representing uplink communication resource constraints and computation capability constraints, betak，nIs the bandwidth allocated to user n by edge server k, fnIs the calculated frequency of user n and
and
is the minimum and maximum of the calculated frequency; c4And C5Ensure the stability and high efficiency of the equipment participating in the model training6Ensuring that the users participating in the training are users within the service range of the edge server.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110958468.1A CN113791895A (en) | 2021-08-20 | 2021-08-20 | Edge calculation and resource optimization method based on federal learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110958468.1A CN113791895A (en) | 2021-08-20 | 2021-08-20 | Edge calculation and resource optimization method based on federal learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113791895A true CN113791895A (en) | 2021-12-14 |
Family
ID=79182029
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110958468.1A Pending CN113791895A (en) | 2021-08-20 | 2021-08-20 | Edge calculation and resource optimization method based on federal learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113791895A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114465900A (en) * | 2022-03-01 | 2022-05-10 | 北京邮电大学 | Data sharing delay optimization method and device based on federal edge learning |
| 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 |
| CN116094993A (en) * | 2022-12-22 | 2023-05-09 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
| CN115983390B (en) * | 2022-12-02 | 2023-09-26 | 上海科技大学 | Edge intelligent reasoning method and system based on multi-antenna aerial calculation |
| CN116827393A (en) * | 2023-06-30 | 2023-09-29 | 南京邮电大学 | Honeycomb-free large-scale MIMO uplink receiving method and system based on federal learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107425895A (en) * | 2017-06-21 | 2017-12-01 | 西安电子科技大学 | A kind of 3D MIMO statistical channel modeling methods based on actual measurement |
| CN111860581A (en) * | 2020-06-11 | 2020-10-30 | 南京邮电大学 | Federal learning training method based on model dispersion |
| CN112351503A (en) * | 2020-11-05 | 2021-02-09 | 大连理工大学 | Task prediction-based multi-unmanned-aerial-vehicle-assisted edge computing resource allocation method |
| US20210073639A1 (en) * | 2018-12-04 | 2021-03-11 | Google Llc | Federated Learning with Adaptive Optimization |
-
2021
- 2021-08-20 CN CN202110958468.1A patent/CN113791895A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107425895A (en) * | 2017-06-21 | 2017-12-01 | 西安电子科技大学 | A kind of 3D MIMO statistical channel modeling methods based on actual measurement |
| US20210073639A1 (en) * | 2018-12-04 | 2021-03-11 | Google Llc | Federated Learning with Adaptive Optimization |
| CN111860581A (en) * | 2020-06-11 | 2020-10-30 | 南京邮电大学 | Federal learning training method based on model dispersion |
| CN112351503A (en) * | 2020-11-05 | 2021-02-09 | 大连理工大学 | Task prediction-based multi-unmanned-aerial-vehicle-assisted edge computing resource allocation method |
Non-Patent Citations (2)
| Title |
|---|
| SIQI LUO 等: "HFEL: Joint Edge Association and Resource Allocation for Cost-Efficient Hierarchical Federated Edge Learning", IEEE * |
| 蔡晓然;莫小鹏;许杰;: "面向联合学习的D2D计算任务卸载", 物联网学报, no. 04 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114465900A (en) * | 2022-03-01 | 2022-05-10 | 北京邮电大学 | Data sharing delay optimization method and device based on federal edge learning |
| 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 |
| CN115983390B (en) * | 2022-12-02 | 2023-09-26 | 上海科技大学 | Edge intelligent reasoning method and system based on multi-antenna aerial calculation |
| CN116094993A (en) * | 2022-12-22 | 2023-05-09 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
| CN116094993B (en) * | 2022-12-22 | 2024-05-31 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
| CN116827393A (en) * | 2023-06-30 | 2023-09-29 | 南京邮电大学 | Honeycomb-free large-scale MIMO uplink receiving method and system based on federal learning |
| CN116827393B (en) * | 2023-06-30 | 2024-05-28 | 南京邮电大学 | Honeycomb-free large-scale MIMO receiving method and system based on federal learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113791895A (en) | Edge calculation and resource optimization method based on federal learning | |
| Liu et al. | Chance-constrained optimization in D2D-based vehicular communication network | |
| CN110401964B (en) | Power control method based on deep learning for user-oriented center network | |
| Chen et al. | A GNN-based supervised learning framework for resource allocation in wireless IoT networks | |
| Sun et al. | Location optimization and user association for unmanned aerial vehicles assisted mobile networks | |
| CN106604300B (en) | Small cell base station self-energy supply and self-return method based on full duplex and large-scale antenna technology | |
| Ji et al. | Trajectory and communication design for cache-enabled UAVs in cellular networks: A deep reinforcement learning approach | |
| CN113490219B (en) | Dynamic resource allocation method for ultra-dense networking | |
| Li et al. | Joint communication and trajectory design for intelligent reflecting surface empowered UAV SWIPT networks | |
| Nouri et al. | Multi-UAV placement and user association in uplink MIMO ultra-dense wireless networks | |
| Wang et al. | Content placement considering the temporal and spatial attributes of content popularity in cache-enabled UAV networks | |
| CN114219354A (en) | Resource allocation optimization method and system based on federal learning | |
| Sun et al. | Location optimization for unmanned aerial vehicles assisted mobile networks | |
| Al-Abiad et al. | Coordinated scheduling and decentralized federated learning using conflict clustering graphs in fog-assisted IoD networks | |
| Nguyen et al. | Optimization of resource allocation for underlay device-to-device communications in cellular networks | |
| Taimoor et al. | Holistic resource management in UAV-assisted wireless networks: An optimization perspective | |
| Zhang et al. | Joint computation offloading and trajectory design for aerial computing | |
| Ji et al. | Reinforcement learning based joint trajectory design and resource allocation for RIS-aided UAV multicast networks | |
| Zhai et al. | Antenna subarray management for hybrid beamforming in millimeter-wave mesh backhaul networks | |
| Fu et al. | Joint speed and bandwidth optimized strategy of UAV-assisted data collection in post-disaster areas | |
| Tang et al. | Optimizing power and rate in cognitive radio networks using improved particle swarm optimization with mutation strategy | |
| CN107249213B (en) | A kind of maximized power distribution method of D2D communication Intermediate Frequency spectrum efficiency | |
| CN115664486A (en) | Energy efficiency optimization method for wireless energy supply in RIS (RIS) assisted UAV (unmanned aerial vehicle) edge computing system | |
| WO2017185994A1 (en) | Clustering method, device and system | |
| Abuzgaia et al. | UAV Communications in 6G Cell-Free Massive MIMO Systems |
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 |