Federated Meta-Learning for task offloading and resource allocation in MEC-IoT

MEC and the Internet of Things (IoT) are two areas rapidly expanding technologies that offer many opportunities to improve efficiency and application performance. However, the huge amount of data generated by IoT devices and the processing and latency constraints imposed by these technologies and mobile networks make processing this data a major challenge. As part of MEC architecture, a promising approach to solving this challenge is to deploy computing servers at the edge of the network, close to IoT devices. This makes it possible to reduce latency and traffic load on the core network, while offering a better user experience. However, offloading tasks from IoT devices to the MEC servers and efficiently allocating the available computing resources is a complex problem. IoT tasks can have latency, bandwidth, and speed requirements, resources, and available computational resources may be limited or shared between multiple users. To solve this problem, we propose an approach based on federated meta-learning. In our approach, each device IoT collects information about the tasks to be performed and the local resources available, then securely shares them with a local MEC server. The local server uses these information to train a meta-learning model that can predict the best task offload and resource allocation decisions for each task. The advantage part of our approach is that it allow us to learn from the experiences of all IoT devices, resulting in more robust and accurate models. We evaluated our FedMeta2Ag algorithm using the MNIST database (Modified National Institute of Standards and Technology database). Considering 20 epochs, the accuracy during training is 91.5% against 92.0% obtained with the test data. In addition, performance continues to increase during the first 20 iterations and gradually becomes stable. Moreover, the accuracy evaluation involves the comparison of the proposed approach with existing methods. This comparison reveals that our approach predict the performance with more accuracy than the existing models. This approach can effectively fulfil the demanding performance of wireless communications systems.

I. Introduction

The concepts of meta-learning and federated learning are not new, but recent advances in gradient-based optimization have brought them back into the spotlight as promising solutions for fast learning. The authors of [1] investigated a task offloading mechanism based on federated reinforcement learning in mobile edge computing. The algorithm this effectiveness were determined by evaluating model loss at different learning rates to determine the change in learning loss. But these algorithms require training from scratch. In order to avoid training from scratch and obtain customized models for heterogeneous data sets, meta-learning is adopted to overcome the statistical heterogeneity problem encountered with federated learning. In particular, Finn and al [2] proposed a gradient-based algorithm called Model-Agnostic Meta-Learning (MAML), which directly optimizes learning performance over model initialization. Thus, the authors of [3] proposed a Meta-Learning-based task offloading algorithm (MELO) that trains a general Deep Neural Network (DNN) for different MEC task scenarios and can quickly learn to adapt to a new task. This algorithm addresses the limitation of federated learning, but requires the transfer of all data, resulting in the disclosure of personal data.