DRL Model for Distributed Agent-based IoT on Multi-Access Edge Computing for Accident Forecast

This paper introduces a state-of-the-art Deep Reinforcement Learning (DRL) model designed for IoT devices in Multi-Access Edge Computing (MEC), with a focus on urban accident prediction. The Edge server serves as a central hub, collecting reinforcement learning data from distributed IoT devices and distributing custom deep learning models for reinforcement learning back to these devices, all within the network. The study also addresses privacy concerns by combining Differential Privacy (DP) with Federated Learning (FL) in DRL within MEC environments. DRL combines deep learning and reinforcement learning principles, enabling wireless IoT devices to optimize actions for maximum rewards in their environments. The surge in IoT devices in networks not only generates massive data but also raises privacy concerns. Differential Privacy, rooted in Google's FL technology, safeguards data privacy. IoT devices utilize DRL and FL to improve learning efficiency and uphold privacy policies. The paper outlines how wireless IoT devices accumulate traceable data tables through reinforcement learning within a Grid-world framework. Within the MEC architecture, Edge servers efficiently execute deep neural networks tailored for reinforcement learning data and distribute results to other IoT devices upon request. This paper presents a tailored DRL model for MEC and a federated DRL framework for maintaining differential privacy while handling shared data generated through deep learning in Edge servers.

I. Introduction

With the exponential proliferation of wireless Internet of Things (IoT) devices [1], the volume of data generated has reached unprecedented levels. As we advance our efforts in developing reinforcement learning techniques for autonomous vehicles, robotics, and gaming, along with deep learning models inspired by biological neural networks, the emergence of deep reinforcement learning (DRL) [2] stands out as a pivotal technology in decentralized distributed networks. DRL represents a fusion of deep learning and reinforcement learning [3], both rooted in the fundamentals of Markov decision processes (MDP) [4] and neural networks. Reinforcement learning entails decision-making at each step of the machine learning process, while deep learning forms the network for decision-making based on the data derived from reinforcement learning. In essence, it allows wireless IoT devices, functioning as autonomous agents, to learn and make decisions through the maximization of rewards obtained from their environment, all without relying on either supervised or unsupervised input data.