Combining Learners to Predict Link Quality in Wireless IoT Networks

In the Internet of Things (IoT) or in industrial networks, the use of poor links in data gathering may considerably degrade network performance. In this paper, we investigate link quality prediction in order to anticipate and avoid the degradation of network performance (i.e. latency, throughput, reliability and lifetime). Several machine learning techniques are investigated to predict the class of any link: Good, poor or Intermediate. All these techniques are based on Random Forests which provides high accuracy, as well as a very good precision and recall per class. They differ in their use of a single link quality metric or a number of metrics. Some rely on a single learner, whereas others build their final decision based on the inputs of several base learners. Their performance is evaluated with per-class criteria as well as global criteria over a data set collected on an actually deployed multichannel network of 50 nodes.