Wireless Mesh Networks Routing Optimization using Machine Learning and Deep Learning

Wireless mesh networks (WMNs) are increasingly deployed in large-scale infrastructures, where efficient and reliable routing is essential to maximise performance. Conventional protocols such as RPL rely on local decision-making and often produce globally suboptimal Directed Acyclic Graphs (DAGs), particularly in dense or complex topologies. This work investigates whether machine learning (ML) and deep learning (DL) techniques can be used to globally optimise DAG-based routing in WMNs. To support this exploration, we developed a high-performance simulator that evaluates the routing efficiency of arbitrary spanning trees using realistic link quality metrics derived from physical parameters such as RSSI. This simulator generates a large labelled dataset of optimal and suboptimal trees across randomly generated topologies, which is then used to train various ML and DL models. We formulate the optimisation task as a multi-label classification problem, where models predict the set of edges forming the optimal DAG. Despite extensive experimentation with classical classifiers, MLPs, and RNNs, our results show that standard models fail to outperform RPL, primarily due to the difficulty of learning under strict DAG constraints, such as acyclicity and single-parent-per-node rules, which render the optimisation problem non-differentiable. However, the simulator consistently identifies DAGs that outperform RPL by up to 16% in terms of routing efficiency. These findings underscore the limitations of unconstrained learning in combinatorial domains and point toward future research into constraint-aware, structure-driven learning frameworks for global WMN optimisation.