A survey on how network simulators serve reinforcement learning in wireless networks

International audienceRapid adoption of mobile devices, coupled with the increase in prominence of mobile applications and services, resulted in unprecedented infrastructure requirements for mobile and wireless networks. To improve user experience, future 5G and wireless network systems evolve to support increased mobile traffic, real-time precision analysis, and adaptable network resource management. As mobile environments become more complex, heterogeneous, and evolving, these tasks become more difficult. In order to solve these problems, many researchers rely on reinforcement learning. The success of reinforcement learning stems from its support for new and powerful tools that solve problems. Nodes mobility, instability of wireless connections, the coexistence of multiple wireless technologies, and resource sharing among users are a few examples of what makes a wireless network a dynamic system. Learning, which is the main feature of reinforcement learning, enables wireless nodes to adapt to the dynamics of the system over time. For the learning to be efficient, it should be done over realistic and varied conditions. This is where network simulation tools can be useful. Network simulators are extensively used when it comes to studying wireless network protocols. They offer the advantage of scaling up scenarios at minimum cost and the ability to test many possible configurations quicker under a controlled environment. The main purpose of this survey is to show how network simulators help in developing reinforcement learning techniques in wireless networks. We emphasize how these tools can be used in the learning process and which problems they can solve. In the end, we discuss open issues related to this topic and highlight some best practice guidelines when it comes to mixing network simulators, reinforcement learning, and wireless protocols