DQN-CAMR: A Deep Q-Network Based Congestion-Aware Multipath Routing Framework for Quality-of-Service Provisioning in Mobile Ad-Hoc Networks

Abstract

Reactive single-path protocols such as AODV commit an entire flow to one route for its lifetime, so when node mobility or bursty traffic pushes queue occupancy at an intermediate node past its buffer capacity, every flow sharing that node suffers correlated delay growth and tail packet loss until a new route is discovered. Multipath extensions such as AOMDV reduce this effect by maintaining backup routes, but typically split traffic using static or hop-count-based ratios that do not react to real-time queue state, leaving the underlying congestion-avoidance problem largely unsolved. This paper formulates per-flow traffic distribution across multiple discovered paths as a constrained Markov Decision Process whose objective is to minimize a weighted combination of end-to-end delay, packet loss, and maximum link utilization, and proposes DQN-CAMR, a Deep Q-Network based Congestion-Aware Multipath Routing framework that learns, from local queue-occupancy, link-quality, and hop-count observations, how to split each flow's traffic across up to three concurrently maintained paths. Congestion is estimated directly from MAC-layer queue occupancy rather than inferred indirectly from delay, allowing the agent to react before queue overflow causes loss rather than after. DQN-CAMR extends AOMDV route-discovery with a lightweight congestion-metric field and replaces its static split ratio with a compact feed-forward Q-network re-evaluated every route-maintenance interval. Simulated in NS-3 across 50-150 nodes and 2-10 Mbps offered load against AOMDV, AODV, and an MDQ-inspired congestion-aware baseline adapted from SDN literature, DQN-CAMR sustains 15-30% higher aggregate throughput at high load, cuts average end-to-end delay by 18-32% at 20-25 concurrent flows, reduces packet loss ratio by roughly half relative to AOMDV at 10 Mbps offered load, and maintains a Jain's fairness index above 0.92 as flow count grows to 25, indicating that congestion-aware, learning-based traffic splitting is a practical route to QoS-aware routing in mobile ad-hoc networks without requiring centralized control.