Context-driven Access Point Selection for Connected Vehicles using Reinforcement Learning

Wifi and small cell networks can relieve the strain on cellular networks, delaying the need for more spectrum, enabling less radiated power, and more energy-efficient communication with mobile nodes. The challenge, however, is to manage the association of mobile nodes to the access point (AP) that helps maintain connectivity. In this research, we propose an approach that integrates context and network information and uses reinforcement learning (RL) to learn AP switching for a vehicle in motion. We examine the importance of context information by comparing agents trained with different types of context: position, mobility and wireless environment. Additionally, we employ a supervised learning approach to allow interpretation of the trained models. The models were trained and evaluated using real-world data collected in an urban environment. Our best RL model outperforms switching based on maximum received signal strength indicator (RSSI), by 38% on one test set and 44% on another. Additionally, the RL trained models have been open-sourced to promote transparency and enable broader adoption in real-world scenarios.