Reinforcement Learning Approaches for IoT Networks with Energy Harvesting

In this paper, we consider power allocation optimization of IoT user with energy harvesting techniques in Internet of Things (IoT) network. We consider an IoT user as an agent, which makes optimal decisions on its transmit power in a time frame. The throughput maximization problem is formulated with the objective function as the cumulative long-term throughput and under the constraints of dynamic energy harvesting , task transmission and channel uncertainty. Here, the environment dynamics are modelled as first-order Markov Decision Processes(MDP), which are generally solved by the Dynamic Programming(DP) method. However, DP method has to know the explicit model of environment dynamics. Thus we propose a near optimal algorithm based on ε-greedy Q-learning to solve this problem with the channel gain, task queue and the available power as the network states. Simulations show that the power allocation with energy harvesting is achieved by both the DP method and the ε-greedy Q-learning algorithm, moreover, the latter algorithm has better throughput performance.