摘要：Unmanned aerial vehicles (UAVs) are supposed to be used to provide different services from video surveillance to communication facilities during critical and high-demanding scenarios. Augmented reality streaming services are especially demanding in terms of required throughput, computing resources at the user device, as well as user data collection for advanced applications, for example, location-based or interactive ones. This work is focused on the experimental utilization of a framework adopting reinforcement learning (RL) approaches to define the paths crossed by UAVs in delivering resources for augmented reality services. We develop an OpenAI Gym-based simulator that is tuned and tested to study the behavior of UAVs trained with RL to fly around a given area and serve augmented reality users. We provide abstractions for the environment, the UAVs, the users, and their requests. A reward function is then defined to encompass several quality-of-experience parameters. We train our agents and observe how they behave as a function of the number of UAVs and users at different hours of the day.
