PROJECT TITLE :

Learning to Solve Task-Optimized Group Search for Social Internet of Things

ABSTRACT:

Because the Internet of Things (IoT) has become more established and widespread, the concept of the Social Internet of Things (SIoT) has emerged as a way to facilitate the development of innovative applications and networking services for the IoT in a manner that is both more effective and efficient. Despite the fact that there are many works for SIoT, their primary focus is on the design of architectures and protocols for SIoT in accordance with particular schemes. There is still a lot of uncharted territory when it comes to figuring out how to collaborate effectively using SIoT to complete difficult tasks. As a result, we propose a new problem family called Task-Optimized SIoT Selection (TOSS), which aims to identify the best group of Internet of Things (IoT) objects to perform a specific set of tasks from a pool of available tasks. The objective of TOSS is to select the target SIoT group in such a way that members of the target SIoT group are able to easily communicate with one another and perform the assigned tasks to the highest possible degree of precision. We demonstrate that both the Bounded Communication-loss TOSS (BC-TOSS) and the Robustness Guaranteed TOSS (RG-TOSS) problems are both NP-hard and inapproximable by presenting two different problem formulations for different scenarios and naming them respectively. We present a polynomial-time algorithm for BC-TOSS that comes with a performance guarantee, as well as a polynomial-time algorithm that is both efficient and effective at obtaining good solutions for RG-TOSS. We further propose Structure-Aware Reinforcement Learning (SARL) to leverage the Graph Convolutional Networks (GCN) and Deep Reinforcement Learning (DRL) to effectively solve RG-TOSS. This is because RG-TOSS is NP-hard and inapproximable within any factor. In addition, because the problem instances that are simulated for DRL using graph models are not the same as the real ones, we propose Structure-Aware Meta Reinforcement Learning (SAMRL) as a method for rapidly adapting to new domains. Our proposed algorithms have been shown to perform better than other deterministic and learning-based baseline approaches, according to the results of experiments conducted on a variety of real-world datasets.