CMML: Combined metaheuristic-machine learning for adaptable routing in clustered wireless sensor networks

Abstract

Cluster-based routing is the most common routing approach to achieve energy efficiency in wireless sensor networks. However, optimal determination of cluster heads is NP-hard, which calls for heuristics or metaheuristics for obtaining a near-optimal solution. Although metaheuristics achieve better performance, they suffer from high computational time, and thus, cannot rapidly respond to routing requests. Also, a large majority of the existing routing protocols cannot easily adapt to changing network or application configurations. In this paper, a Combined model based on Metaheuristics and Machine Learning, named CMML, is proposed to support efficient and adaptable routing in clustered wireless sensor networks. In our CMML model, a multi-criteria heuristic clustering algorithm is used for clustering in which a metaheuristic (e.g., genetic algorithm) is utilized for the automatic tuning of the heuristic algorithm for each configuration separately. We repeat this process for several configurations (i.e., for different network sizes, numbers of nodes, aggregation factors, lifetime definitions, etc.). The tuned heuristic algorithm in each configuration is subsequently used for network simulation to obtain the corresponding solution. As a result, a comprehensive dataset for different configurations is derived, which is used to train a machine learning model (e.g., support vector machine). The input feature vector of a sample comprises local features (current state of a node at a round), global features (current state of the network), and application-specific features, while the output is the priority factor of each node to be selected as a cluster head. After training the CMML model, it can be applied as a quickly adaptable clustering protocol. In fact, our motivation is to utilize the generalizability of machine learning to learn the behavioral pattern of the metaheuristic algorithm in finding best routes for previous configurations. Simulation results demonstrate that the CMML model can effectively adapt with different applications, while prolonging the network lifetime based on the application requirements.