IoT and autonomous systems are in charge of an increasing number of sensing, processing and communications tasks. These systems may be equipped with energy harvesting devices. Nevertheless, the energy harvested is uncertain and variable, which makes it difficult to manage the energy
in these systems. Reinforcement learning algorithms can handle such uncertainties, however selecting the adapted algorithm is a difficult problem. Many algorithms are available and each has its own advantages and drawbacks. In this paper, we try to provide an overview of different approaches
to help designer to determine the most appropriate algorithm according to its application and system. We focus on Q-learning, a popular reinforcement learning algorithm and several of these variants. The approach of Q-learning is based on the use of look up table, however some algorithms use
a neural network approach. We compare different variants of Q-learning for the energy management of a sensor node. We show that depending on the desired performance and the constraints inherent in the application of the node, the appropriate approach changes.
Reinforcement-Learning Approach Guidelines for Energy Management
