论文笔记——Federated learning framework for mobile edge computing networks
This paper focuses on the application of federal learning in demand forecasting.
Generally speaking, FL has some problems:
- Non-independent and identically distributed data. Customer training data sets vary, and a given local training data set does not represent population distribution.
- Unbalanced data set. Each client has a different amount of local training data. This means that different customers have different reliability of training values, because the training procedure may be too short.
- Mass distribution. The client involved in the FL framework is significantly larger than the number of data processed locally.
- Limited communication. Mobile devices are not always available for training data, and often they may be slow or have poor communication conditions.
This paper proposes a hybrid cloud MEC network solution, which considers a virtual machine copy VRC that deploys most of the requesting applications on the network element NE located at the edge of the network ED and admitted to the edge equipment.
Appropriate VRC allocation is based on the hit rate AHP, which indicates the hit rate when the search device requests the application of the group on the NE near it.
The main contribution of this research is
- Applying the contextualization of the FL architecture to the deployment of VRCs, using decentralized training data is a powerful tool for finding effective results on the distribution of VRCs
- Without exhausting the hardware resources of EDs, the basic method of performing distributed data training is detected and applied.
- A large number of numerical simulations have been carried out and compared with the method of chaos theory (CT) to verify the remarkable performance of applying this method to the deployment of VRCs.
Network structure:
Each ED that requires task calculation mainly finds the VRC of a desired application on a nearby NEs. (We assume that each ED sends the task to the nearest NE, if it does not contain the application requested by the ED, then forward the task to the nearest NE that contains the application. Finally, we assume that each NE knows that other NEs contain VRCs, and each NE stores a routing table, which stores the shortest path between these NEs for each pair of NEs.) If no NE has the VRC of the required application, the task is sent to all application types There are clouds.
The total calculation cost (OCC) experienced by ED j when performing calculations is determined by
Given,
Where 
and 
represent the time that task j spends on NE i and the cloud, respectively. It should be noted that 
and 
are expressed as the sum of the task execution time spent in the CPUs of the NE and the cloud, and the queuing time for the task to wait for execution at these sites. (Assume that the CPU queue uses a first-in first-out service strategy) In addition, it 
indicates the transmission cost when sending a task from ED j to its nearest NE i *. If task j is calculated on NEi, then 
the binary value of is 1, otherwise it is 0.
algorithm:
The proposed FL framework consists of a client and a server (usually a base station) responsible for local data training. The base station is represented by a central server designed to improve the global learning model by merging and aggregating local models updated by EDs. This method is based on the interaction process between the client and the server. During each iteration of the algorithm, the EDs involved in the training process are a subset of the entire EDs set, and the number of elements is equal to y.
The data set used in the experiment is MovieLens 1M data set, hereinafter referred to as data set 1, and MovieLens 100K data set [34] is referred to as data set 2 in this paper.
In order to verify the effectiveness of the proposed method, the accuracy of the predicted values obtained by the CT method, the phase space reconstruction method and the proposed deep learning-based method were compared.
