Intelligent subflow steering in MPTCP-based hybrid Wi-Fi and LiFi networks using model-augmented DRL

A hybrid Wi-Fi and light fidelity (LiFi) network combines the best of two worlds with the ubiquitous coverage of Wi-Fi and the high peak data rate of LiFi as the radio spectrum does not interfere with the light spectrum. This hybrid network might be realized by using multipath TCP (MPTCP), where Wi-Fi and LiFi paths can be simultaneously employed to potentially boost the total throughput of Wi-Fi while increasing the resilience towards network failure of LiFi due to, for example, blockage. However, naively implementing MPTCP in a hybrid Wi-Fi and LiFi network can yield an unexpected result, such as a lower throughput compared to the single-path TCP due to a Head-of-Line delay during the slow start phase of the TCP congestion control. Even though this problem can be avoided by improving the existing flow control or congestion control of TCP, these solutions still lack intelligent decision making that can improve the adaptability of MPTCP. Therefore, in this paper, we propose a model-augmented deep reinforcement learning (DRL) approach to intelligently steer MPTCP subflows (i.e., TCP connections) by using a close-to-reality scenario emulated by considering random orientation, random blockage, and random mobility of Wi-Fi-and-LiFi-enabled mobile devices. As a result, we will show later that a performance gain can be achieved compared to the state-of-the-art while maintaining ease implementation to existing MPTCP implementations