S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Dept. of Computer Science and Engineering (컴퓨터공학부) Theses (Master's Degree_컴퓨터공학부)
A Wireless Network Selection Scheme based on Machine Learning
기계 학습 기반 무선망 선택 기법
- 공과대학 컴퓨터공학부
- Issue Date
- 서울대학교 대학원
- 학위논문 (석사)-- 서울대학교 대학원 : 컴퓨터공학부, 2017. 2. 김종권.
- Today, more and more public Wi-Fi APs are deployed for user convenience in response to the growing number of smartphone users. However, when both Wi-Fi and LTE interface are on, mobile devices usually select Wi-Fi interface regardless of its connection quality or stability. The blind preference of Wi-Fi interface can degrade user experience called QoE (Quality of Experience).
We present a decision tree based network connection management named SISA (Smart Interface Selection Algorithm) which takes account of user context, contents and network features such as RSSI, Link Speed, TCP throughput and Gateway RTT (RTT to gateway) to efficiently select wireless radio interfaces. We observed the effect of intermittent connection or continuous connection with low quality Wi-Fi AP on the TCP throughput. Through extensive measurements and experiments in real-field, especially public places, we confirmed that the mobile devices embedded our platform outperforms commodity devices for various scenarios. We also discovered that there is no correlation between network features measured in real-field.
We extend the problem to MPTCP (Multipath TCP) that allows the devices like smartphones and tablets to exploit both interfaces concurrently. Similar with phenomenon over SPTCP (Single Path TCP), in real mobile environments where wireless devices abound and quality of active paths changes frequently, it makes Wi-Fi connection affect the MPTCP performance negatively. To solve the problem, we present a novel path management called MPTCP-ML (MPTCP based on Machine Learning). It manages the usage among multiple paths based on decision calculated by machine learning model. We use path quality metrics as inputs for machine learning model. For accurate capturing of path quality, we utilize different quality metrics including signal strength, data rate, TCP throughput, the number of APs on the same and adjacent channel, and RTT (Round Trip Time). We have implemented MPTCP-ML in Android and conducted experiments for various and dynamic mobile environments. The results show that MPTCP-ML outperforms generic MPTCP, especially for mobile environments.