Federated learning method and system for network uncertainty awareness in mobile edge cloud
By defining the average amount of training parameters and model size factors of user equipment in the MEC network, determining the appropriate number and location of the aggregators is solved, and the federated learning FL framework that minimizes costs is realized optimization.
Patent Information
- Application Number
- CN202110629960.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-06-07
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-06-07
AI Technical Summary
In MEC networks, the number of aggregators is unreasonable, resulting in the problem of minimizing implementation costs under uncertainty.
By defining the average amount of training parameters for each user device under the uncertainty of the MEC network, the average model size factor during the federated learning FL task request process is determined, and the minimum and maximum number of aggregators is determined based on this factor, and an auxiliary graph is constructed to determine the decision. Aggregator location and quantity, optimize the federated learning FL framework to minimize total cost.
The optimization of aggregator placement and user equipment allocation in the MEC network is realized, reducing the total cost in the federated learning FL task request process and improving resource utilization efficiency.
Smart Images

Figure CN113255934B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of MEC network communications, and in particular to a federated learning method and system for network uncertainty perception in a mobile edge cloud. Background Art
[0002] With the rapid development of 5G and Mobile Edge Computing (MEC), various AI applications such as augmented reality and smart healthcare have been deployed in MEC networks. Such AI applications continuously generate a large amount of data in MEC networks, which needs to be analyzed to improve its accuracy. Traditional training is performed in a centralized location (i.e., data center) with a powerful graphics processor by sending all data from each user equipment (UE) to the centralized location. However, sending all raw data (including sensitive data, i.e., user facial images) to a centralized location may lead to user privacy violations. Federated Learning (FL) is considered a promising technology to avoid privacy leakage of user devices by training the generated data in user devices. The training parameters in each user device are then sent to a centralized location to be merged with other user devices to obtain a global machine learning model. Since the data for AI applications is generated at the edge, MEC is a natural technology for implementing cost-sensitive federated learning.
[0003] Enabling FL in MEC networks poses fundamental challenges due to fully distributed computing resources and network uncertainty. Specifically, 5G base stations are densely deployed to ensure full coverage of user devices as they adopt mmWave for communication. First, such dense deployment brings dynamicity and network uncertainty to a new level, which makes the implementation of FL requests more complicated. User devices can register to different base stations in a short period of time because the coverage area of 5G base stations is small. This means that the arrival pattern of FL requests is uncertain. In addition, if each user device has available resources at this time, intermittent local training is performed. In addition, before implementing its FL request, the parameter size of each FL service may be uncertain because its neural network model is unknown. This uncertainty complicates the placement of aggregators. If the aggregator is placed after the request arrives, in a certain MEC network, the overhead will get out of control due to latency or other reasons, and it is impractical to place aggregators for consecutive FL requests of different models. Therefore, when the parameter size is uncertain, it is very important to place aggregators in advance in order to be able to process incoming requests in a timely manner. Second, the resource capacity of base stations in dense 5G MEC networks requires that aggregators be carefully placed in multiple edge locations. Specifically, limited computing resources, such as floating-point op amps, neural network accelerators, are attached to 5G base stations. Therefore, forward link services may not fit into a single base station. The performance of this layout depends on communication and processing costs, as the trained models need to be sent out for aggregation. Third, implementing FL services requires not only careful placement of aggregators, but also finding the right number of aggregators. If fewer aggregators are used in FL services, processing costs can be saved. However, this may result in high communication costs because users send their trained models to aggregators over longer paths. Therefore, finding the right number of aggregators for each FL service is a challenge. Summary of the invention
[0004] The purpose of the present invention is to provide a network uncertainty-aware federated learning method and system in a mobile edge cloud to solve the problem that the number of aggregators in the existing MEC network is unreasonable and the implementation cost cannot be minimized under the uncertainty of the MEC network.
[0005] To achieve the above object, the present invention provides the following solutions:
[0006] A network uncertainty-aware federated learning method in a mobile edge cloud, comprising:
[0007] Based on the federated learning FL framework, the average amount of training parameters of each user device under MEC network uncertainty is defined; the MEC network uncertainty is the uncertainty of the transmitted model parameters;
[0008] Determine an average model size factor in each federated learning FL task request process according to the average amount of training parameters of each user device;
[0009] Determine the minimum number of aggregators and the maximum number of aggregators under each federated learning FL task request process according to the average model size factor;
[0010] Determining the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators;
[0011] Building an auxiliary graph according to the number of aggregators, and determining a location decision according to the auxiliary graph; the location decision includes an allocation location of a user device, a placement location of an aggregator, and a placement location of a service;
[0012] Based on the location decision, determining a total cost incurred in each of the federated learning FL task request processes according to the aggregate quantity;
[0013] Taking the resource capacity of the MEC network as a constraint, adjusting the number of aggregators according to the total cost, generating the optimal number of aggregators in each federated learning FL task request process, and optimizing the federated learning FL framework according to the optimal number of aggregators to minimize the total cost.
[0014] Optionally, determining the average model size factor in each federated learning FL task request process according to the average amount of training parameters of each user device specifically includes:
[0015] According to the average amount of the training parameters of each user equipment, discretize the range of the average model size factor into any interval of fixed length;
[0016] determining a finite set of values for the average model size factor based on the fixed length;
[0017] Determining an active value set of the average model size factor according to the finite value set using a greedy algorithm of a multi-armed bandit machine;
[0018] The average model size factor is determined based on the set of activity values.
[0019] Optionally, determining the minimum number of aggregators and the maximum number of aggregators under each federated learning FL task request process according to the average model size factor specifically includes:
[0020] use Determine the minimum number of aggregators; where n min is the minimum number of aggregators; γ q The amount of computing resources allocated for aggregating a unit of data; χ is the average model size factor; μ(|w m|) is the average amount of training parameters for each user device, w m is the size of the transmission model between the user equipment and the FL service m of the user equipment; με m The set of user devices serving FL m; Loc q is a first potential location of a micro cloud or base station; C q For the location Loc q The computing resource capacity of q The amount of bandwidth resources allocated for transmitting unit data; B q is the bandwidth resource capacity;
[0021] Obtain the number of user devices; the number of user devices is equal to the maximum number of aggregators.
[0022] Optionally, determining the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators specifically includes:
[0023] The number of aggregators in the current range is determined according to the minimum number of aggregators and the maximum number of aggregators by using a binary search method; the current range is a range that changes continuously when the binary search method is used.
[0024] Optionally, constructing an auxiliary graph according to the number of aggregators, and determining a location decision according to the auxiliary graph specifically includes:
[0025] constructing an auxiliary graph according to the number of aggregators, and setting the cost and capacity of each edge in the auxiliary graph;
[0026] Considering the training parameters of each user device as commodities with demand, and determining the commodity demand of the training parameters in each federated learning FL request according to the average model size factor;
[0027] Based on the commodity demand, determining a separable flow from a message source to a sink node for each user device from the auxiliary graph; the separable flow is a multi-commodity flow;
[0028] Determine the probability of assigning the user equipment to the base station according to the splittable flow, and place the aggregator in the Loc q The probability of service placement to Loc q The probability of Loc q A first potential location for a micro cloud or base station;
[0029] According to the probability of the user equipment being assigned to the base station, the aggregator is placed at the Loc q The probability that the service is placed in Loc q The probability of user equipment randomly assigned to the base station and the aggregator randomly placed as Locq 's location and where services are randomly placed;
[0030] According to the position of the user equipment randomly assigned to the base station, in the user equipment and base station layer of the auxiliary graph, each splittable flow of the user equipment is moved to a randomly selected base station;
[0031] According to the aggregator, it is randomly placed as Loc q , in the aggregation layer of the auxiliary graph, moving each splittable flow of the user equipment to the aggregator with the minimum cost;
[0032] According to the randomly placed location of the service, in the service layer of the auxiliary graph, each splittable flow of the user equipment is moved to the location where the service is located;
[0033] Determine the indivisible flow based on the separable flow and the divisible flow, and convert the indivisible flow into a location decision including the allocation location of the user device, the placement location of the aggregator, and the placement location of the service; the node path passed by the indivisible flow is the decision result of the location decision.
[0034] Optionally, the probability of the user equipment being allocated to the base station is:
[0035]
[0036] Among them, p k,i The probability of assigning a user equipment to a base station; f k (bs i ) is the traffic passing through the edge between the user equipment and the base station in the user equipment and base station layer of the auxiliary graph; f k traffic obtained for the user equipment;
[0037] In the aggregator layer, there is a potential location for the aggregator for FL requests; create n m Parts W m,o , each component corresponds to an aggregator A m,o The potential locations with sufficient available resources are added to the component W m,o In order to realize the aggregator A m,o Complete the aggregation task; create the second potential location Loc' q , the second potential location Loc' q is a virtual location node, Loc' q With Loc q Add to the components together;
[0038] The aggregator is placed in Loc q The probability is:
[0039]
[0040] Among them, p m,o,q Placed to Loc for aggregator q The probability of k is the kth user equipment; UE is the general term for user equipment; W m,o For aggregator A m,o Parts of Loc q' is the second potential location; BS is the user equipment and base station layer; CL is the micro cloud; f k (Loc q , W m,o ) is the edge through the auxiliary graph <Loc q , Loc' q >Component W m,o The traffic of the route; Through aggregator A m,o Place in a Loc q Position component W m,o The total routing traffic, Through aggregator A m,o Place in all Loc q Position component W m,o The total routing traffic.
[0041] At the service layer, for each first potential location Loc q and Service m Create two virtual nodes Loc″ q and Loc″′ q , Loc″ q is the third potential location, Loc″′ q For the fourth potential position; every two Loc" q and Loc″′ q are added to a new component for service S m ;
[0042] The service is placed in Loc q The probability is:
[0043]
[0044] Among them, p m,q Place the service in Loc q The probability of k (Loc q , S m ) is the edge through the auxiliary graph <Loc″ q , Loc″′ q >Routed traffic; For service S m Place in a Loc q Total routing traffic for a location, For service S m Place in all Loc q Total routing traffic for the location.
[0045] Optionally, the location-based decision and determining the total cost incurred in each federated learning FL task request process according to the aggregate quantity specifically include:
[0046] Using the formula Determine the total cost incurred in each of the federated learning FL task request processes; wherein cost(n) is the total cost incurred in the federated learning FL task request process using n aggregators, and n is the number of aggregations; The computational cost of training the dataset locally on the user device; For service S m User equipment ue k Upload to an aggregator A m,o communication costs; For service S m The communication cost of uploading the model from an aggregator to the main aggregator; is the cost of aggregating parameters during the federated learning FL task request process.
[0047] A network uncertainty-aware federated learning system in a mobile edge cloud, comprising:
[0048] The average amount definition module of the training parameters is used to define the average amount of the training parameters of each user equipment under the uncertainty of the MEC network based on the federated learning FL framework; the uncertainty of the MEC network is the uncertainty of the model parameters transmitted;
[0049] An average model size factor determination module, used to determine an average model size factor in each federated learning FL task request process according to an average amount of training parameters of each user device;
[0050] A minimum aggregator number and a maximum aggregator number determination module, used to determine the minimum aggregator number and the maximum aggregator number under each federated learning FL task request process according to the average model size factor;
[0051] an aggregator quantity determination module, configured to determine the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators;
[0052] A location decision determination module, configured to construct an auxiliary graph according to the number of aggregators, and determine a location decision according to the auxiliary graph; the location decision includes an allocation location of a user device, a placement location of an aggregator, and a placement location of a service;
[0053] A total cost determination module, configured to determine a total cost generated in each federated learning FL task request process based on the location decision and the aggregate quantity;
[0054] An adjustment module is used to adjust the number of aggregators according to the total cost based on the resource capacity of the MEC network as a constraint, generate an optimal number of aggregators in each federated learning FL task request process, and optimize the federated learning FL framework according to the optimal number of aggregators to minimize the total cost.
[0055] Optionally, the average model size factor determination module specifically includes:
[0056] A discretization unit, used for discretizing the range of the average model size factor into any interval of a fixed length according to the average amount of the training parameters of each user equipment;
[0057] a finite value set determining unit, configured to determine a finite value set of the average model size factor according to the fixed length;
[0058] an active value set determining unit, configured to determine the active value set of the average model size factor according to the finite value set by using a greedy algorithm of a multi-armed bandit machine;
[0059] The average model size factor determining unit is used to determine the average model size factor according to the activity value set.
[0060] Optionally, the minimum number of aggregators and the maximum number of aggregators determining module specifically include:
[0061] Minimum number of aggregators to determine the unit for utilizing Determine the minimum number of aggregators; where n min is the minimum number of aggregators; γ q The amount of computing resources allocated for aggregating a unit of data; χ is the average model size factor; μ(|w m |) is the average amount of training parameters for each user device, w m is the size of the transmission model between the user equipment and the FL service m of the user equipment; με m The set of user devices serving FL m; Loc q is a first potential location of a micro cloud or base station; C q For the location Loc qThe computing resource capacity of q The amount of bandwidth resources allocated for transmitting unit data; B q is the bandwidth resource capacity;
[0062] The maximum aggregator number determining unit is used to obtain the number of user devices; the number of user devices is equal to the maximum aggregator number.
[0063] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the present invention provides a network uncertainty-aware federated learning method and system in a mobile edge cloud. Aiming at the aggregator placement and user device allocation problems in a single FL task request process in an MEC network, based on a federated learning FL framework, the average amount of training parameters of each user device under MEC network uncertainty is defined, the average model size factor in each federated learning FL task request process is determined, and the appropriate number of aggregators is determined according to the average model size factor, an auxiliary graph is constructed, and the positions and numbers of aggregators are reasonably arranged based on the auxiliary graph, the number of aggregators is adjusted with the resource capacity of the MEC network as a constraint, the optimal number of aggregators in each federated learning FL task request process is generated, and the federated learning FL framework is optimized according to the optimal number of aggregators to minimize the total cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0065] Figure 1 A flow chart of the federated learning method for network uncertainty perception in a mobile edge cloud provided by the present invention;
[0066] Figure 2 This is a schematic diagram of the federated learning FL framework provided by the present invention;
[0067] Figure 3 Auxiliary diagram provided by the present invention;
[0068] Figure 4 This is a structural diagram of the network uncertainty-aware federated learning system in the mobile edge cloud provided by the present invention. DETAILED DESCRIPTION
[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0070] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0071] Figure 1 The flowchart of the federated learning method for network uncertainty perception in the mobile edge cloud provided by the present invention is as follows: Figure 1 As shown, a network uncertainty-aware federated learning method in a mobile edge cloud includes:
[0072] Step 101: Based on the federated learning FL framework, define the average amount of training parameters of each user equipment under MEC network uncertainty; the MEC network uncertainty is the uncertainty of the transmitted model parameters.
[0073] Step 102: Determine an average model size factor in each federated learning FL task request process according to the average amount of training parameters of each user device.
[0074] In practical applications, step 102 specifically includes: discretizing the range of the average model size factor into any interval of a fixed length according to the average amount of training parameters of each user device; determining a finite value set of the average model size factor according to the fixed length; using a greedy algorithm of a multi-armed bandit to determine an active value set of the average model size factor according to the finite value set; and determining the average model size factor according to the active value set.
[0075] Step 103: Determine the minimum number of aggregators and the maximum number of aggregators under each federated learning FL task request process according to the average model size factor.
[0076] In practical applications, the step 103 specifically includes: using Determine the minimum number of aggregators; where n min is the minimum number of aggregators; γ q The amount of computing resources allocated for aggregating a unit of data; χ is the average model size factor; μ(|w m |) is the average amount of training parameters for each user device, w m is the size of the transmission model between the user equipment and the FL service m of the user equipment; με mThe set of user devices serving FL m; Loc q is a first potential location of a micro cloud or base station; C q For the location Loc q The computing resource capacity of q The amount of bandwidth resources allocated for transmitting unit data; B q is the bandwidth resource capacity;
[0077] Obtain the number of user devices; the number of user devices is equal to the maximum number of aggregators.
[0078] Step 104: Determine the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators.
[0079] In practical applications, the step 104 specifically includes: using a binary search method to determine the number of aggregators in the current range according to the minimum number of aggregators and the maximum number of aggregators; the current range is a range that changes continuously during the binary search method.
[0080] Step 105: construct an auxiliary graph according to the number of aggregators, and determine a location decision according to the auxiliary graph; the location decision includes the allocation location of the user equipment, the placement location of the aggregator, and the placement location of the service.
[0081] In practical applications, the step 105 specifically includes: constructing an auxiliary graph according to the number of aggregators, and setting the cost and capacity of each edge in the auxiliary graph; treating the training parameters of each user device as a commodity with demand, and determining the commodity demand of the training parameters in each federated learning FL request according to the average model size factor; based on the commodity demand, determining a splittable flow from a message source to a sink node for each user device from the auxiliary graph; the splittable flow is a multi-commodity flow; determining the probability of the user device being assigned to a base station and the aggregator being placed to a Loc according to the splittable flow q The probability of placing the service at Loc q The probability of Loc q is a first potential location of a micro cloud or a base station; according to the probability of the user equipment being assigned to the base station, the aggregator is placed at Loc q The probability that the service is placed in Loc q The probability of user equipment randomly assigned to the base station and the aggregator randomly placed as Loc q The location of the user equipment and the location where the service is randomly placed; according to the location of the user equipment randomly assigned to the base station, in the user equipment and base station layer of the auxiliary graph, each splittable flow of the user equipment is moved to a randomly selected base station; according to the location of the aggregator randomly placed as Loc qAccording to the position of the service randomly placed, in the aggregation layer of the auxiliary graph, each divisible flow of the user equipment is moved to the aggregator with the lowest cost; according to the position of the service randomly placed, in the service layer of the auxiliary graph, each divisible flow of the user equipment is moved to the position where the service is located; according to the divisible flow and the divisible flow, the indivisible flow is determined, and the indivisible flow is converted into a location decision including the allocation position of the user equipment, the placement position of the aggregator and the placement position of the service; the node path passed by the indivisible flow is the decision result of the location decision.
[0082] The probability of the user equipment being allocated to the base station is:
[0083]
[0084] Among them, p k,i The probability of assigning a user equipment to a base station; f k (bs i ) is the traffic passing through the edge between the user equipment and the base station in the user equipment and base station layer of the auxiliary graph; f k The traffic obtained for the user equipment.
[0085] In the aggregator layer, there is a potential location for the aggregator for FL requests; create n m Parts W m,o , each component corresponds to an aggregator A m,o The potential locations with sufficient available resources are added to the component W m,o In order to realize the aggregator A m,o Complete the aggregation task; create the second potential location Loc' q , the second potential location Loc' q is a virtual location node, Loc' q With Loc q Add to the assembly together.
[0086] The aggregator is placed in Loc q The probability is:
[0087]
[0088] Among them, p m,o,q Placed to Loc for aggregator q The probability of k is any user equipment; UE is a general term for user equipment; W m,o For aggregator A m,o Parts of Loc q' is the second potential location; BS is the user equipment and base station layer; CL is the micro cloud; f k (Locq , W m,o ) is the edge through the auxiliary graph <Loc q , Loc' q >Component W m,o The traffic of the route; Through aggregator A m,o Place in a Loc q Position component W m,o The total routing traffic, Through aggregator A m,o Place in all Loc q Position component W m,o The total routing traffic.
[0089] At the service layer, for each first potential location Loc q and Service m Create two virtual nodes Loc″ q and Loc″′ q , Loc″ q is the third potential location, Loc″′ q For the fourth potential position; every two Loc" q and Loc″′ q are added to a new component for service S m .
[0090] The service is placed in Loc q The probability is:
[0091]
[0092] Among them, p m,q Place the service in Loc q The probability of k (Loc q , S m ) is the edge through the auxiliary graph <Loc″ q , Loc″′ q >Routed traffic; For service S m Place in a Loc q Total routing traffic for a location, For service S m Place in all Loc q Total routing traffic for the location.
[0093] Step 106: Based on the location decision, determine the total cost generated in each federated learning FL task request process according to the aggregated quantity.
[0094] In practical applications, the step 106 specifically includes: using the formula Determine the total cost incurred in each of the federated learning FL task request processes; wherein cost(n) is the total cost incurred in the federated learning FL task request process using n aggregators, and n is the number of aggregations; The computational cost of training the dataset locally on the user device; For service S m User equipment ue k Upload to an aggregator A m,o communication costs; For service S m The communication cost of uploading the model from an aggregator to the main aggregator; is the cost of aggregating parameters during the federated learning FL task request process.
[0095] Step 107: Taking the resource capacity of the MEC network as a constraint, adjusting the number of aggregators according to the total cost, generating the optimal number of aggregators in each federated learning FL task request process, and optimizing the federated learning FL framework according to the optimal number of aggregators to minimize the total cost.
[0096] Figure 2 This is a schematic diagram of the federated learning FL framework provided by the present invention, such as Figure 2 As shown in the figure, the federated learning (FL) framework is a hierarchical structure consisting of a master aggregator and multiple aggregators that collect trained models from user equipment (UE). A federated learning service usually has multiple rounds of training and parameter aggregation. In each round of the federated learning service, each user equipment trains its local data based on the model parameters and sends the updated model parameters to its aggregator. Then, each aggregator collects the trained parameters from the user equipment. The master aggregator aggregates the data from all aggregators to obtain the global model, and sends the updated global model to all user equipment for the next round of local training.
[0097] Based on the network uncertainty-aware federated learning method in the mobile edge cloud provided by the present invention, the present invention adopts three algorithms to implement the above technical solution, and the three algorithms are: FWK algorithm, ApproAG algorithm and HeuAG algorithm.
[0098] Among them, the FWK algorithm assumes that the model parameter size of each FL request is |w m | is given, and the number of aggregators is obtained using the binary search method. The UE allocation and aggregator placement obtained using the ApproAG algorithm are further used to obtain the cost, and the number of aggregators to be calculated is further changed using the cost.
[0099] The ApproAG algorithm is an offline algorithm that obtains the number of aggregators based on the FWK algorithm when the size of the transmitted model parameters is known, and makes decisions such as UE allocation and aggregator placement.
[0100] HeuAG algorithm: When the size of the transmitted model parameters is uncertain, that is, when the network is uncertain, first estimate the value of a model parameter size, use an online algorithm based on a multi-armed bandit machine and call the FWK algorithm to implement each arriving FL request.
[0101] The technical solution of the present invention is to determine the average model size factor and the minimum number of aggregators and the maximum number of aggregators by averaging the training parameters of each user device under the uncertainty of the MEC network defined by the HeuAG algorithm; the determined minimum number of aggregators and the maximum number of aggregators are used as the known minimum number of aggregators and the maximum number of aggregators in the FWK algorithm, the aggregation number is obtained by using the FWK algorithm, and the ApproAG algorithm is called to determine the location decision according to the aggregation number; the FWK algorithm is called to determine the total cost generated in each federated learning FL task request process according to the aggregation number based on the location decision; finally, the resource capacity of the MEC network is used as a constraint condition, the number of aggregators is adjusted according to the total cost, the optimal number of aggregators in each federated learning FL task request process is generated, and the federated learning FL framework is optimized according to the optimal number of aggregators to minimize the total cost.
[0102] The following specifically introduces three algorithms used in the present invention:
[0103] 1) FWK algorithm: An optimization framework for finding the appropriate number of aggregators for each federated learning task request.
[0104] The number of aggregators in one FL request plays a crucial role in minimizing the implementation cost. More specifically, for one FL request, a larger number of aggregators enables the aggregators to be distributed closer to the user device, thereby reducing the cost of updating the training parameters from the user device to the aggregator. However, this may increase the amount of data transmitted from the aggregator to the server in one FL request, resulting in communication costs between the aggregator and the server.
[0105] In order to find the right number of aggregators for each FL request, a binary search method is used. Obviously, each FL request requires at least one aggregator. The maximum number of aggregators that can be instantiated in a mobile edge network (MEC) depends on the available resources at each location. Consider a 5G mobile edge network G = (BS ∪ CL, E), which consists of a group of small base stations BS and a group of micro clouds CL. Let n min and n max are the minimum and maximum number of aggregators per FL request, nmin The minimum value of is 1, which means that all user device training data are aggregated by a single aggregator, but it is possible that the available resources at this location are insufficient to aggregate these parameter data. Define n min The values are:
[0106]
[0107] Among them, γ q The amount of computing resources allocated to aggregate unit data volume, w m is the size of the model transmitted between the user device and its FL service m, με m The set of user devices serving m for FL, δ q The amount of bandwidth resources allocated for transmitting unit data, B q is the bandwidth resource capacity, Loc q A potential location for a micro cloud or base station.
[0108] Considering the worst case, each aggregator aggregates the data of a single UE, that is, n max The values are:
[0109] n max =|με m | (2)
[0110] Consider an array of elements, each element is [n min , n max ] and sorted in increasing order of their values. and are the minimum and maximum values of the current range, cost(n) is the cost of the FL process using n aggregators, initially, The current range is the range that keeps changing during the binary search.
[0111] [n min , n max ] is obtained as follows:
[0112] definition For user equipment ue k The computing cost consumed by training unit data is the user equipment ue k Local training dataset ds k The computational cost for:
[0113]
[0114] definition For user equipment ue k Through the base station bs i To location Locq The cost of transmitting a unit of data, x m,o,q Represents aggregator A m,o Is it placed in Loc q The binary variable, z k,i Indicates user equipment ue k Whether through base station bs i Access to the MEC network, then in service S m User equipment ue k Upload to an aggregator A m,o Communication cost for:
[0115]
[0116] Define y m,q For Service m Whether it is placed in Loc q A binary variable, Indicates that from the location Loc q To location Loc' q The cost of transmitting a unit of data. m The communication cost of uploading the model from an aggregator to the main aggregator for:
[0117]
[0118] Definition H m Request r for deployed FL m The set of aggregators, α q Indicates at location Loc q Aggregate the training parameter cost of UE, then for FL request r m Cost of Aggregation Parameters for:
[0119]
[0120] Therefore, the cost cost(n) of the FL process using n aggregators is:
[0121]
[0122] Based on this, the optimization framework process for finding the appropriate number of aggregators for each federated learning task request is as follows:
[0123] According to formulas (1) and (2), set n min and n max The value of .
[0124] when hour:
[0125]
[0126] Call the ApproAG algorithm to obtain the aggregator location placement and user allocation decisions and substitute them into formula (3) to obtain cost (n m ).
[0127] calculate and Until
[0128] if but otherwise,
[0129] 2) ApproAG algorithm: Joint aggregator placement and user device allocation algorithm in federated learning tasks.
[0130] The joint aggregator placement and UE allocation problem of a single FL request in MEC network is simplified to a multi-commodity flow problem with minimum cost. That is, the training parameters of each UE are regarded as a commodity, including the source UE k , aggregation node s and demand size |w m |. According to the number of aggregators obtained from the FWK framework, we can construct Figure 3 Auxiliary diagram of With the number of aggregators, we can consider the placement of aggregators and the related user device allocation issues. Zhong Congue k Routing to s, in order to route the UE's goods to their sink node s, A divisible multi-commodity flow from each UE to s is found in . Let f k It is user equipment ue k The obtained flow, f k Can be decomposed into auxiliary graphs Then, the obtained stream is converted into the user equipment ue using the Randomized rounding method. k The probability of being associated with a base station, aggregator placement, and service requested by each FL.
[0131] First calculate the user equipment ue k With base station bs i Associated probability: At the BS layer, use f k (bs i ) indicates that the edge and Traffic of user equipment ue k With base station bs i The associated probability p k,ifor:
[0132]
[0133] Then calculate an aggregator A m,o Place to Loc q Probability: Each A m,o has a widget consisting of its potential position, defining W m,o It is an aggregator A m,o The widget, f k (Loc q , W m,o ) indicates that the edge <Loc q , Loc' q >Widget W m,o Aggregator A m,o Place to Loc q The probability p m,o,q for:
[0134]
[0135] The idea behind this is that the more traffic is routed through a widget edge, the more likely it is to be selected as an A. m,o Third, according to the auxiliary diagram The number of flows routed on each edge of the service layer (Service Layer) is calculated to serve S m Place at location Loc q The probability of . Similar to the probability calculation of the aggregator layer (AggregatorLayer), use f k (Loc q , S m ) indicates that the edge <Loc″ q , Loc″′ q >Route traffic, |w max | represents the maximum value of the transmission model size, d unit Expressed as the ratio of the transmission model size |w m | Smaller data unit. Then service S m Place at location Loc q The probability p m,q for:
[0136]
[0137] Based on this, the steps of the approximation algorithm ApproAG are as follows:
[0138] Constructing auxiliary graphs And set the cost and capacity of each edge accordingly.
[0139] The training parameters of each UE are considered as commodities with demand.
[0140] From the auxiliary graph constructed For each UE, find the source UE k Splittable flows to sink node s.
[0141] According to formulas (4), (5), and (6), the probability p of assigning a user equipment to a base station is calculated respectively: k,i , the aggregator is placed at Loc q The probability of position p m,o,q , the service is placed in Loc q The probability of position p m,q .
[0142] According to the required p k,i Probability of UEue k Randomly assigned to base station bs i According to p m,o,q The probability that each n m Aggregators are randomly placed at Loc q Location, according to p m,q The probability will serve S m Randomly placed in Loc q Location.
[0143] Based on the preliminary FL training framework built earlier, for example, there are selected UEs and corresponding aggregators, where the aggregators are placed, for each user equipment ue k Perform the following steps to find an optimal UE allocation and aggregator placement:
[0144] In the auxiliary diagram In the user equipment and base station layers, f k Each split flow moves to a randomly selected base station.
[0145] In the auxiliary diagram In the polymerization layer, f k Each split flow moves to a placed aggregator with the lowest cost.
[0146] In the auxiliary diagram In the service layer, f k Each split stream is moved to the location where its service is located.
[0147] The resulting indivisible stream This translates into decisions on UE allocation, aggregator and service placement.
[0148] Indivisible Flow The node path passed by is itself the result of decisions made by UE allocation, aggregator and service placement.
[0149] Steps to apply:
[0150] When calculating the cost in FWK, you must know these three decisions to calculate the UE cost (consumption of computing resources and communication resources), the aggregator cost, etc. For example, only after the UE is allocated can the corresponding cost be calculated according to the allocated UE. Then, the number of aggregators is obtained based on the cost using the binary search method.
[0151] Only after making these three decisions can we further complete a complete federated learning task request under resource constraints.
[0152] 3) HeuAG algorithm: An online algorithm that implements each arriving federated learning task request under network uncertainty.
[0153] Since the amount of data in the user device and the size of the learned model parameters are uncertain, it is fundamentally challenging to achieve accurate FL at the edge of the network. This algorithm is a new online learning algorithm based on the multi-armed bandit (MAB) technology to solve the problem of uncertainty in the amount of data in the user device and the size of the learned parameters. The basic idea is to use the optimization framework proposed by FWK, given the model size of each FL request. In FWK, it is assumed that the model size of each FL request|w m | is given. If the model size is uncertain, a feasible approach is to implement each FL request based on the average model size.
[0154] Define μ(|w m |) For each user device ue k The average amount of training parameters, so the FWK framework can be slightly modified, first according to μ(|w m |) to calculate the minimum number of aggregators n min Instead of |w m |, then consider the demand for each commodity in ApproAG as |w m |. However, the actual model size is much larger than the average size, which will seriously exceed the resource capacity of a certain location in the MEC network. To avoid this serious resource conflict, the average model size of each FL request is enlarged by a factor represented by χ, where χ is in [1, |w max | / μ(|w m |)] is the true value. Specifically, n in FWK is calculated by the following method min :
[0155]
[0156] It should be noted that formula (1) is to calculate n by knowing the parameter value of the model size in advance. min , while (7) is under network uncertainty, that is, the model size is estimated by the factor represented by χ.
[0157] In addition, it is considered that the user equipment ue requested by FL k Each commodity has a demand size like this:
[0158] (1+χ)·μ(|w m |) / d unit (8)
[0159] A key question is how to determine the value of χ, taking a fixed value for χ and performing a one-time optimization for requests entering the system one by one. That is, because the distribution of the model sizes of the requests is known, a one-time optimization can be performed to pre-calculate the probabilities of the ApproAG algorithm. When each request arrives, user device allocation, aggregator and service placement are performed randomly according to the pre-calculated probabilities. However, if the actual model size of each request deviates significantly from the expected model size, the actual performance of the algorithm may differ greatly from the optimal performance. To avoid this, a custom scaling algorithm is used to adaptively learn the value of χ when requests arrive in the system.
[0160] In order to find an approximate value of χ more effectively, the range of χ is discretized into ζ intervals, whose length is fixed to len = (|w max | / |w min |-1) / ζ, thus obtaining a finite set of values for χ: val = {1+(i-1)·len|1≤i≤ζ}. Adaptively find a value in val, and by adopting an activation and selection scheme that maintains a set of active values for χ, the minimum cost of the problem can be achieved. Define val` as the set of active values for χ, and adopt an activation and selection scheme for active values, the purpose is to achieve the minimum cost of the problem. Obviously val is the set of finite values of the discretization of χ.
[0161] Once a value in val is activated, it is put into the set val′. Observe that slightly different values of χ are effective in implementing FL requests r m , may incur similar costs. Activating these values with similar costs will reduce the efficiency of the algorithm. Therefore, a confidence radius is defined for each value v∈val`. The idea is to activate the value v` in val to the set val` instead of the confidence radius of v. For a value v and the currently arriving request r m , using num a (v) indicates a request r m Choose v as the number of requests before the value of χ, and use u c(v) indicates the number of times the algorithm selects v. a (v) and denote the number of requests by R. Therefore, the confidence radius is defined as
[0162] Consider each value in val` as an arm, the arm radius is the confidence radius of the value, and the arm in val` is the representative of the value in val. Ideally, the arm radius of val` should cover all values in val. Initially, the coverage radius of each value is large because it has not been selected yet. As requests arrive at the system, the confidence domain of each value is shrinking. This will leave some values uncovered. In order to keep val`'s arms covering all values in val, these uncovered values are added directly to val`. After obtaining a set of active arms, an arm will be selected from val` when each FL request arrives. Ideally, it is hoped to select an arm that not only reduces the cost but also contains more values in val. Therefore, the arms in val` are reduced by 1 / u c (v)+2·radius v Arrange in descending order. The ultimate goal is to select these arms and each arm has an arm radius that can cover all values in val. For example: there are a total of 10 points (equivalent to the set of finite values val), and 3 points are selected from them (equivalent to the set of active values val`) which are equivalent to arms, and then each arm has an arm length (equivalent to drawing a circle with the arm radius as the radius with this point as the center), and these three circles can contain all 10 points.
[0163] Purpose: To achieve FL request r with slightly different χ values (estimated model size) m , which may generate similar costs (equivalent to selecting values from val one by one, that is, trying all 10 points one by one, which will generate 10 additional costs). Activating these values with similar costs will reduce the efficiency of the algorithm. Therefore, the activation and selection scheme of the active value (i.e., val`, that is, only looking for 3 points from it) is adopted to achieve the minimum cost of the problem.
[0164] Based on this, the process of the HeuAG algorithm is as follows:
[0165] The range of the average model size χ transmitted during the FL process is discretized into ζ intervals of fixed length.
[0166] Let val be a finite set of values of χ.
[0167] / *The value of val is selected from val* / .
[0168] Using the greedy approach of a multi-armed bandit, greedily select some values in val that can cover all values in val by radius and add them to val`.
[0169] For each incoming FL request, the following steps are performed:
[0170] Select a highest rank 1 / u from val` c (v)+2·radius v The value of .
[0171] Call the FWK algorithm and set n according to equation (7) min , set the demand for each product according to equation (8); where the value of val` is the estimated average model size μ|w m |, for the case of position transfer model size, after selecting a value from val′, n can be calculated min , the demand for each product (that is, the model parameter size of different UEs | w m |), and implement the FL request by calling FWK.
[0172] Updates the selection count for each value in val` by the confidence radius for each value in val`.
[0173] If there are values (the values of val) not in val`, then: add these uncovered values to val`.
[0174] Figure 4 The structure diagram of the network uncertainty-aware federated learning system in the mobile edge cloud provided by the present invention is as follows: Figure 4 As shown, a network uncertainty-aware federated learning system in a mobile edge cloud includes:
[0175] The average amount definition module 401 of training parameters is used to define the average amount of training parameters of each user equipment under MEC network uncertainty based on the federated learning FL framework; the MEC network uncertainty is the uncertainty of the transmitted model parameters.
[0176] The average model size factor determination module 402 is used to determine the average model size factor in each federated learning FL task request process according to the average amount of training parameters of each user equipment.
[0177] The average model size factor determination module 402 specifically includes: a discrete unit, used to discretize the range of the average model size factor into any interval of a fixed length according to the average amount of the training parameters of each user device; a finite value set determination unit, used to determine the finite value set of the average model size factor according to the fixed length; an active value set determination unit, used to determine the active value set of the average model size factor according to the finite value set using the greedy algorithm of the multi-armed bandit machine; and an average model size factor determination unit, used to determine the average model size factor according to the active value set.
[0178] The minimum aggregator number and maximum aggregator number determination module 403 is used to determine the minimum aggregator number and the maximum aggregator number under each federated learning FL task request process according to the average model size factor.
[0179] The minimum number of aggregators and the maximum number of aggregators determining module 403 specifically include: a minimum number of aggregators determining unit, configured to use Determine the minimum number of aggregators; where n min is the minimum number of aggregators; γ q The amount of computing resources allocated for aggregating a unit of data; χ is the average model size factor; μ(|w m |) is the average amount of training parameters for each user device, w m is the size of the transmission model between the user equipment and the FL service m of the user equipment; με m The set of user devices serving FL m; Loc q is a first potential location of a micro cloud or base station; C q For the location Loc q The computing resource capacity of q The amount of bandwidth resources allocated for transmitting unit data; B q is the bandwidth resource capacity; a maximum aggregator number determination unit is used to obtain the number of user equipment; the number of user equipment is equal to the maximum aggregator number.
[0180] The aggregator quantity determination module 404 is used to determine the aggregator quantity according to the minimum aggregator quantity and the maximum aggregator quantity.
[0181] The location decision determination module 405 is used to construct an auxiliary graph according to the number of aggregators and determine a location decision according to the auxiliary graph; the location decision includes the allocation location of the user equipment, the placement location of the aggregator and the placement location of the service.
[0182] The total cost determination module 406 is configured to determine the total cost generated in each federated learning FL task request process based on the location decision and the aggregated quantity.
[0183] The adjustment module 407 is used to adjust the number of aggregators according to the total cost based on the resource capacity of the MEC network as a constraint, generate the optimal number of aggregators in each federated learning FL task request process, and optimize the federated learning FL framework according to the optimal number of aggregators to minimize the total cost.
[0184] The present invention aims at the joint aggregator placement and user device allocation problem of a single FL process in an MEC network, taking the resource capacity of the MEC network as a constraint to minimize the implementation cost; the present invention proposes an optimization framework to find the appropriate number of aggregators for each FL request, and proposes an approximation algorithm, which has a provable approximation ratio for the defined problem; the present invention considers the uncertainty of the amount of data of each user device and the size of the learned parameters of each user device request, and considers the user device allocation problem and the online aggregator placement problem in the MEC network of multiple FL requests. For this purpose, a new online learning algorithm is proposed using the multi-armed bandit (MAB) technology; the performance of the proposed algorithm is evaluated based on existing research. Experimental results show that the performance of the proposed algorithm is better than that of similar algorithms, and the implementation cost is reduced by at least 15%.
[0185] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0186] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A network uncertainty-aware federated learning approach in mobile edge clouds, It is characterized in that include: Based on the federated learning FL framework, the average amount of training parameters for each user device under uncertainty in the MEC network is defined; The MEC network uncertainty refers to the uncertainty of the model parameters transmitted; Determine an average model size factor in each federated learning FL task request process according to the average amount of training parameters of each user device; Determine the minimum number of aggregators and the maximum number of aggregators under each federated learning FL task request process according to the average model size factor; Determining the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators; An auxiliary graph is constructed according to the number of aggregators, and a location decision is determined according to the auxiliary graph, specifically including: constructing an auxiliary graph according to the number of aggregators, and setting the cost and capacity of each edge in the auxiliary graph; Considering the training parameters of each user device as commodities with demand, and determining the commodity demand of the training parameters in each federated learning FL request according to the average model size factor; Based on the commodity demand, determining a separable flow from a message source to a sink node for each user device from the auxiliary graph; the separable flow is a multi-commodity flow; Determine the probability of assigning the user equipment to the base station according to the splittable flow, and place the aggregator to The probability of placing the service probability; A first potential location for a micro cloud or base station; According to the probability of the user equipment being assigned to the base station, the aggregator is placed The probability that the service is placed The probability of random assignment of user equipment to the base station and random placement of aggregators is determined by The location of the service and where it is randomly placed; According to the position of the user equipment randomly assigned to the base station, in the user equipment and base station layer of the auxiliary graph, each splittable flow of the user equipment is moved to a randomly selected base station; According to the aggregator, it is randomly placed as , in the aggregation layer of the auxiliary graph, moving each splittable flow of the user equipment to the aggregator with the minimum cost; According to the randomly placed location of the service, in the service layer of the auxiliary graph, each splittable flow of the user equipment is moved to the location where the service is located; Determine the indivisible flow according to the separable flow and the divisible flow, and convert the indivisible flow into a location decision including the allocation location of the user equipment, the placement location of the aggregator, and the placement location of the service; the node path passed by the indivisible flow is the decision result of the location decision; the location decision includes the allocation location of the user equipment, the placement location of the aggregator, and the placement location of the service; The probability of the user equipment being allocated to the base station is: in, probability of assigning a user equipment to a base station; The traffic passing through the edge between the user equipment and the base station in the user equipment and base station layer of the auxiliary graph; traffic obtained for the user equipment; In the Aggregator Layer, there is a potential location for the aggregator for FL requests; create Parts , each component corresponds to an aggregator potential locations with sufficient available resources are added to the component To implement aggregator Complete aggregation task; create second potential location , the second potential location is a virtual location node, and Add to the components together; The aggregator is placed The probability is: in, Place it for the aggregator probability; Any user equipment; UE is a general term for user equipment; For Aggregators Components; is the second potential location; BS is the user equipment and base station layer; CL is the micro cloud; is the edge through the auxiliary graph Parts The traffic of the route; For aggregators Place in a Position Widget The total routing traffic, For aggregators Place in all Position Widget Total routing traffic; At the service level, for each first potential location and Services Create two virtual nodes and , is the third potential position, For the fourth potential position; every two and are added to a new component for service ; The service is placed in The probability is: in, Place for service probability; is the edge through the auxiliary graph Routed traffic; For service through Place in a Total routing traffic for a location, For service through Place in all Total routing traffic for a location; | w max | Indicates the maximum size of the transmission model. d unit Expressed as a ratio of the transmission model size |w m | Smaller data units; Based on the location decision, determining a total cost incurred in each of the federated learning FL task request processes according to the aggregate quantity; Taking the resource capacity of the MEC network as a constraint, adjusting the number of aggregators according to the total cost, generating the optimal number of aggregators in each federated learning FL task request process, and optimizing the federated learning FL framework according to the optimal number of aggregators to minimize the total cost.
2. The method for network uncertainty awareness federated learning in mobile edge cloud according to claim 1, It is characterized in that The determining of the average model size factor in each federated learning FL task request process according to the average amount of training parameters of each user device specifically includes: According to the average amount of the training parameters of each user equipment, discretize the range of the average model size factor into any interval of fixed length; determining a finite set of values for the average model size factor based on the fixed length; Determining an active value set of the average model size factor according to the finite value set using a greedy algorithm of a multi-armed bandit machine; The average model size factor is determined based on the set of activity values.
3. The method for network uncertainty awareness federated learning in mobile edge cloud according to claim 1, It is characterized in that The determining, according to the average model size factor, the minimum number of aggregators and the maximum number of aggregators under each federated learning FL task request process specifically includes: use Determine the minimum number of aggregators; where, is the minimum number of aggregators; The amount of computing resources allocated to aggregate a unit of data volume; is the average model size factor; is the average amount of training parameters for each user device, For the FL service of the user equipment and the user equipment The size of the model transferred between them; Serving FL A collection of user devices; A first potential location for a micro cloud or base station; C q For location Computing resource capacity; The amount of bandwidth resources allocated for transmitting unit data; is the bandwidth resource capacity; Obtain the number of user devices; the number of user devices is equal to the maximum number of aggregators.
4. The method for federated learning of network uncertainty awareness in mobile edge cloud according to claim 1, It is characterized in that The determining the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators specifically includes: The number of aggregators in the current range is determined according to the minimum number of aggregators and the maximum number of aggregators by using a binary search method; the current range is a range that changes continuously when the binary search method is used.
5. The method for network uncertainty awareness federated learning in mobile edge cloud according to claim 1, It is characterized in that The determining, based on the location decision and according to the aggregate quantity, of the total cost generated in each federated learning FL task request process specifically includes: Using the formula Determine the total cost incurred during each of the federated learning FL task requests; wherein, For use The total cost incurred in the process of requesting a federated learning FL task for each aggregator, where n is the number of aggregations; The computational cost of training the dataset locally on the user device; For service Medium User Equipment Upload to an aggregator communication costs; For service The communication cost of uploading the model from an aggregator to the main aggregator; is the cost of aggregating parameters during the federated learning FL task request process.
6. A network uncertainty-aware federated learning system in mobile edge clouds, It is characterized in that include: The module for defining the average amount of training parameters is used to define the average amount of training parameters of each user device under uncertainty in the MEC network based on the federated learning FL framework; The MEC network uncertainty refers to the uncertainty of the model parameters transmitted; An average model size factor determination module, used to determine an average model size factor in each federated learning FL task request process according to an average amount of training parameters of each user device; A minimum aggregator number and a maximum aggregator number determination module, used to determine the minimum aggregator number and the maximum aggregator number under each federated learning FL task request process according to the average model size factor; an aggregator quantity determination module, configured to determine the number of aggregators according to the minimum number of aggregators and the maximum number of aggregators; A location decision determination module, configured to construct an auxiliary graph according to the number of aggregators, and determine a location decision according to the auxiliary graph; the location decision includes an allocation location of a user device, a placement location of an aggregator, and a placement location of a service; An auxiliary graph is constructed according to the number of aggregators, and a location decision is determined according to the auxiliary graph, specifically including: constructing an auxiliary graph according to the number of aggregators, and setting the cost and capacity of each edge in the auxiliary graph; Considering the training parameters of each user device as commodities with demand, and determining the commodity demand of the training parameters in each federated learning FL request according to the average model size factor; Based on the commodity demand, determining a separable flow from a message source to a sink node for each user device from the auxiliary graph; the separable flow is a multi-commodity flow; Determine the probability of assigning the user equipment to the base station according to the splittable flow, and place the aggregator to The probability of placing the service probability; A first potential location for a micro cloud or base station; According to the probability of the user equipment being assigned to the base station, the aggregator is placed The probability that the service is placed The probability of random assignment of user equipment to the base station and random placement of aggregators is determined by The location of the service and where it is randomly placed; According to the position of the user equipment randomly assigned to the base station, in the user equipment and base station layer of the auxiliary graph, each splittable flow of the user equipment is moved to a randomly selected base station; According to the aggregator, it is randomly placed as , in the aggregation layer of the auxiliary graph, moving each splittable flow of the user equipment to the aggregator with the minimum cost; According to the random placement of the service, each splittable flow of the user equipment is moved to the location of the service in the service layer of the auxiliary graph. Place; Determine the indivisible flow according to the separable flow and the divisible flow, and convert the indivisible flow into a location decision including the allocation location of the user equipment, the placement location of the aggregator, and the placement location of the service; the node path passed by the indivisible flow is the decision result of the location decision; The probability of the user equipment being allocated to the base station is: in, probability of assigning a user equipment to a base station; The traffic passing through the edge between the user equipment and the base station in the user equipment and base station layer of the auxiliary graph; traffic obtained for the user equipment; In the Aggregator Layer, there is a potential location for the aggregator for FL requests; create Parts , each component corresponds to an aggregator potential locations with sufficient available resources are added to the component To implement aggregator Complete aggregation task; create second potential location , the second potential location is a virtual location node, and Add to the components together; The aggregator is placed The probability is: in, Place it for the aggregator probability; Any user equipment; UE is a general term for user equipment; For Aggregators Components; is the second potential location; BS is the user equipment and base station layer; CL is the micro cloud; is the edge through the auxiliary graph Parts The traffic of the route; For aggregators Place in a Position Widget The total routing traffic, For aggregators Place in all Position Widget Total routing traffic; At the service level, for each first potential location and Services Create two virtual nodes and , is the third potential position, For the fourth potential position; every two and are added to a new component for service ; The service is placed in The probability is: in, Place for service probability; is the edge through the auxiliary graph Routed traffic; For service through Place in a Total routing traffic for a location, For service through Place in all Total routing traffic for a location; | w max | Indicates the maximum size of the transmission model. d unit Expressed as a ratio of the transmission model size |w m |Smaller data units; A total cost determination module, configured to determine a total cost generated in each federated learning FL task request process based on the location decision and the aggregate quantity; An adjustment module is used to adjust the number of aggregators according to the total cost based on the resource capacity of the MEC network as a constraint, generate an optimal number of aggregators in each federated learning FL task request process, and optimize the federated learning FL framework according to the optimal number of aggregators to minimize the total cost.
7. The network uncertainty-aware federated learning system in mobile edge cloud according to claim 6, It is characterized in that The average model size factor determination module specifically includes: A discretization unit, used for discretizing the range of the average model size factor into any interval of a fixed length according to the average amount of the training parameters of each user equipment; a finite value set determining unit, configured to determine a finite value set of the average model size factor according to the fixed length; an active value set determining unit, configured to determine the active value set of the average model size factor according to the finite value set by using a greedy algorithm of a multi-armed bandit machine; The average model size factor determining unit is used to determine the average model size factor according to the activity value set.
8. The network uncertainty-aware federated learning system in mobile edge cloud according to claim 6, It is characterized in that The minimum number of aggregators and the maximum number of aggregators determining module specifically include: Minimum number of aggregators to determine the unit for utilizing Determine the minimum number of aggregators; where, is the minimum number of aggregators; The amount of computing resources allocated to aggregate a unit of data volume; is the average model size factor; is the average amount of training parameters for each user device, For the FL service of the user equipment and the user equipment The size of the model transferred between them; Serving FL A collection of user devices; A first potential location for a micro cloud or base station; C q For location Computing resource capacity; The amount of bandwidth resources allocated for transmitting unit data; is the bandwidth resource capacity; The maximum aggregator number determining unit is used to obtain the number of user equipment; the number of user equipment is equal to the maximum aggregator number.
