Explainable Neural Network-Driven Context-Aware Predictive Routing for QoS Optimization in MANETs

Abstract

– Mobile Ad Hoc Networks (MANETs) play a critical role in disaster recovery, military communications, vehicular networking, emergency response, and remote monitoring applications. However, sustaining Quality of Service (QoS) in MANETs is challenging because of changing topologies, mobile nodes, energy constraints, unreliable wireless links, and congestion. Widely adopted routing protocols, including Ad hoc On-Demand Distance Vector (AODV), Dynamic Source Routing (DSR), Destination-Sequenced Distance Vector (DSDV), and Optimized Link State Routing (OLSR), rely on reactive or proactive route discovery mechanisms and often fail to adapt efficiently to rapidly changing network conditions. Although Deep Reinforcement Learning (DRL)-based routing approaches have demonstrated improved adaptability, they suffer from convergence instability, high computational complexity, continuous online learning requirements, and limited interpretability, restricting their practical deployment in mission-critical environments. To overcome these challenges, this paper presents an Explainable Neural Network-Driven Context-Aware Predictive Routing (XNN-CPR) method. The proposed method predicts the reliability of network links, selects routes based on QoS requirements, and uses SHAP to explain routing decisions. All these functions are combined into a single routing framework. The proposed method uses contextual network parameters, which include node mobility, residual energy, queue occupancy, Received Signal Strength Indicator (RSSI), and Signal-to-Noise Ratio (SNR), to infer future link reliability and congestion conditions. An explainable neural network is trained using simulation-generated data to estimate route stability, while SHAP-based explanations are employed to interpret and validate routing decisions. Unlike conventional DRL-based routing approaches that require continuous policy optimization during network operation, the proposed framework separates model training from routing execution, thereby reducing computational overhead and improving decision stability. Simulation experiments conducted in the NS-3 environment show that the proposed XNN-CPR framework steadily outperforms AODV, DSR, OLSR, and Deep Deterministic Policy Gradient (DDPG)- based routing protocols. Compared with the DRL-based benchmark, XNN-CPR obtains a 3.94% improvement in Packet Delivery Ratio (PDR), a 17.67% reduction in End-to-End Delay, and an 8.69% rise in Throughput. The results make sure that the integration of predictive intelligence with explainable decision-making noticeably enhances routing reliability, network stability, and QoS performance while ensuring transparency and trustworthiness in highly dynamic MANET environments.