A Cooperative Multiagent Approach for Optimal Drone Deployment Using Reinforcement Learning

The use of drones in the context of mobile networks has seen an exponential increase in the last couple of years due to the potential advantages in terms of line of sight, mobility, and agility when compared to traditional ground networks. Despite these advantages, the optimal deployment of aerial networks is still challenging, with several works proposing different implementation alternatives. In this work, we propose a Q-learning algorithm based on a cooperative multiagent approach to intelligently find the optimal positions for a set of drones. The main objective of the proposed solution is to minimize the number of users in outage, depending on the density of users, through the dynamic assignment of transmission frequencies and considering whether each drone is participating in the communication stage. We propose and compare four different strategies for the Q-learning algorithm with different action selection policies, whose algorithms differ in terms of design complexity, ability to vary the number of drones in operation, and convergence time. The results show that as the density of users in the region of interest increases, the number of frequencies in operation must increase. In addition, for a single frequency the ALL strategy obtains the best results in all scenarios. But for three and six frequencies the New strategy gets the best results in all scenarios.