RLProph: a dynamic programming based reinforcement learning approach for optimal routing in opportunistic loT networks

摘要

Routing in Opportunistic Internet of Things networks (OppIoTs) is a challenging task because of intermittent connectivity between devices and the lack of a fixed path between the source and destination of messages. Recently, machine learning (ML) and reinforcement learning (RL) have been used with great success to automate processes in a number of different problem domains. In this paper, we seek to fully automate the OppIoT routing process by using the Policy Iteration algorithm to maximize the possibility of message delivery. Moreover, we model the OppIoT environment as a Markov decision process (MDP) replete with states, actions, rewards, and transition probabilities. The proposed routing protocol, RLProph, is able to optimize the routing process via the optimal policy obtained by solving the MDP using Policy Iteration. Through extensive simulations, we show that RLProph outperforms a number of ML-based and context-aware routing protocols on a multitude of performance criteria.