A deep reinforcement learning-based multi-optimality routing scheme for dynamic IoT networks

With the development of Internet of Things (IoT) and 5G technologies, more and more applications, such as autonomous vehicles and tele-medicine, become more sensitive to network latency and accuracy, which require routing schemes to be more flexible and efficient. In order to meet such urgent need, learning-based routing strategies are emerging as strong candidate solutions, with the advantages of high flexibility and accuracy. These strategies can be divided into two categories, centralized and distributed, enjoying the advantages of high precision and high efficiency, respectively. However, routing becomes more complex in dynamic IoT network, where the link connections and access states are time-varying, hence these learning-based routing mechanisms are required to have the capability to adapt to network changes in real time. In this paper, we designed and implemented both centralized and distributed Reinforcement Learning-based Routing schemes combined with Multi-optimality routing criteria (RLR-M). By conducting a series of experiments, we performed a comprehensive analysis of the results and arrived at the conclusion that the centralized is better suited to cope with dynamic networks due to its faster reconvergence (2.2 over distributed), while the distributed is better positioned to handle with large-scale networks through its high scalability (1.6 over centralized). Moreover, the multi-optimality routing scheme is implemented through model fusion, which is more flexible than traditional strategies and as such is better placed to meet the needs of IoT.