A Reliable and Low-Latency Graph-Routing Approach for IWSN using Q-Routing

For Industrial Wireless Sensor Networks, the Network Manager is responsible for the overall configuration, route definition, allocation of communication resources, and optimization of the network. Graph routing is used to increase the reliability of the network through path redundancy. Reinforcement Learning models have been used to optimize latency, energy consumption, and data delivery. Q- Routing is a learning model where nodes in the network learn which of its neighboring nodes provide the best routes for a destination node. Despite presenting good results, this model is not applicable to centralized networks since it does not provide path redundancy and nodes are not allowed to choose routes. We present the Q- Learning Reliable Routing with Multiple Agents approach, that builds routing graphs in a centralized manner using Q-Routing. Each node is represented by a learning agent. Periodically, each agent acts by choosing neighbors used to forward data to a destination. An updated graph is then built and configured over the network. Rewards are given to each agent when its average data latency decreases. Simulations were conducted on a WirelessHART simulator. Results show, in most cases, a reduction of the average network latency while the communication reliability is at least as good as the state-of-the-art graph-routing algorithms.