A Machine Learning Approach for IEEE 802.11 Channel Allocation

TLDR: This work focuses on a subset of problems, identifying Wireless Local Area Networks (WLANs) that severely interfere with each other, and proposes a channel allocation that optimises performance and as a side effect, stabilises networks that the authors do not control.

Abstract: Today’s communication is mainly done over wireless networks, with IEEE 802.11 (Wi-Fi) at the forefront. There are billions of devices and millions of access points (APs), but only very few non-overlapping channels. As a result, the performance of Wi-Fi devices is severely degraded, because perfect channel allocation - with every AP alone in its channel - is close to impossible. Even in situations where all networks are under centralised control, existing approaches quickly tend to be either unscalable or suboptimal. By focusing on a subset of problems, identifying Wireless Local Area Networks (WLANs) that severely interfere with each other, performance can be improved even in such a complex situation. We tackle this problem through machine learning and coin it Bad Neighbour Detection (BND). Based on this output alongside monitoring data about the networks’ activity, we then propose a channel allocation that optimises performance and as a side effect, stabilises networks that we do not control. We evaluate our approach in a field trial and show that we significantly improve the experience for users, eliminating virtually all interference-related issues.