Enhancing Variable Neighborhood Search (VNS) Performance for Relay Node Deployment through Machine Learning

Abstract

Wireless Sensor Networks (WSNs) have become over the years a very attractive field of research. In fact, they had the attention of many researchers who have been interested in issues raised by these networks, such as energy, deployment, coverage, connectivity, latency, routing, etc. WSNs are particularly characterized by their miniaturized aspect, which makes them stealthy, and have rapid deployment in ac cessible or inaccessible zones. WSNs, which are considered an emerging technology, have a wide variety of applications in various fields such as military, health, trans portation, agriculture, etc. In this thesis, we address the hybridization of Machine Learning (ML) tools and meta-heuristics to optimize the deployment of Relay Nodes (RNs) and Network Diameter (ND) in WSNs dedicated to the surveillance of sensi tive fenced areas (e.g., oil/nuclear sites, airport, etc.). Consequently, we propose a novel hyper-heuristics, labeled STAT_UCB_GVNS and DYN_UCB_GVNS, that use Reinforcement Learning (RL) to guide the local search process, enhancing the effectiveness of the VNS algorithm in terms of solution quality. Experimental re sults demonstrate that our RL-based approach, achieves significant improvements in fitness value, specifically in the number of deployed RNs and ND, compared to Basic VNS (BVNS). Indeed, the proposed RL-based approach achieves an average fitness improvement of 49.97% while deploying an average of 53.144% fewer relays than BVNS.