Using machine learning techniques to make packet delivery more reliable in opportunistic networks
Résumé
Opportunistic Networks (OppNets), considered as a subclass of Delay Tolerant Networks (DTNs), provide an effective communication solution for environments characterized by intermittent connectivity and highly dynamic topology. Their operation relies on the store-carry-forward principle, whereby mobile nodes temporarily store packets and forward them whenever a communication opportunity arises. Although this paradigm enables communication without fixed infrastructure, it remains limited by major challenges such as high latency, congestion, link variability, unpredictable node mobility, and energy constraints. To improve the reliability and performance of such networks, this Ph.D. thesis proposes an innovative approach based on the integration of Machine Learning (ML) techniques to make OppNets more intelligent, adaptive, and autonomous. Three main contributions are developed. First, a routing model based on Reinforcement Learning (RL) is designed to enable nodes to dynamically learn the most efficient forwarding strategies according to the network state and past experiences. Simulation results. show that this model improves the packet delivery ratio (PDR) by 35%, reduces the average delay by 40%, and decreases energy consumption by 25% compared to the classical epidemic routing protocol. Second, a Supervised Learning (SL) approach using algorithms such as Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) is developed to predict the probability of successful transmission and thus reduce redundant forwarding. Among these models, Random Forest achieves the best accuracy of 91%, outperforming SVM (76.6%) and MLP (83.3%), while reducing redundant transmissions by 40% and increasing throughput by 53.3%. Third, an Unsupervised Learning (UL) approach based on K-Means, DBSCAN, and Agglomerative algorithms is implemented to perform intelligent clustering of nodes exhibiting similar mobility and contact patterns. Experimental results reveal that the K-Means algorithm offers the best trade-off between computational cost and clustering quality, achieving a silhouette index of 0.45 and reducing forwarding energy consumption by 3.2mj. Altogether, these approaches demonstrate an overall network performance improvement with an average increase in delivery rate of 21%, a latency reduction of 12.4ms, and an energy efficiency improvement of 4.9 mj. The major scientific contribution of this thesis lies in demonstrating that integrating ML techniques into OppNets transforms traditional static and probabilistic routing protocols into self-learning systems capable of dynamically adapting to changing network conditions. This approach introduces intelligent and context-aware decision mechanisms that optimize resource usage, reduce congestion, and ensure more reliable packet delivery. The results of this research open new perspectives for the design of next-generation opportunistic networks that are more robust, autonomous, and efficient, applicable to vehicular networks, the Internet of Things (IoT), sensor systems, and emergency communications.
Citer ce document
Exporter : BibTeX · RIS (Zotero, Mendeley, EndNote)
Accès au document
Voir sur le dépôt sourceCe document est hébergé sur son dépôt institutionnel d'origine.
Licence et provenance
Licence : CC BY
Notice moissonnée depuis HAL le 20/10/2026. Le document reste hébergé par sa source.
Voir le document à la source →