An adaptive multipath routing algorithm based on deep reinforcement learning with delayed-reward in dynamic IoT networks

Routing optimization in dynamic networks remains a challenging problem due to the complexity of network topologies and the non-stationary nature of traffic flows. Traditional routing protocols such as RIP and ECMP often fail to fully utilize network resources under high traffic loads, while conventional reinforcement learning methods struggle to converge in complex environments. To overcome these limitations, this paper proposes a Deep Reinforcement Learning with Delayed Reward for Multi-Path Routing (MPRDRL) algorithm. In MPRDRL, each router node is modeled as an independent agent that learns to make routing decisions based on local observations. Unlike prior reinforcement learning approaches that rely on instantaneous rewards, our method introduces a path-based delayed reward mechanism that evaluates the global transmission performance of each packet upon arrival at its destination, allowing agents to more accurately evaluate their actions based on long-term outcomes. Under experimental scenarios, MPRDRL outperforms various baselines in terms of end-to-end delay and maximum link utilization.