Decision Tree and Random Forest Implementations for Fast Filtering of Sensor Data

With increasing capabilities of energy efficient systems, computational technology can be deployed, virtually everywhere. Machine learning has proven a valuable tool for extracting meaningful information from measured data and forms one of the basic building blocks of ubiquitous computing. In high-throughput applications, measurements are rapidly taken to monitor physical processes. This brings modern communication technologies to its limits. Therefore, only a subset of measurements, the interesting ones, should be further processed and possibly communicated to other devices. In this paper, we investigate architectural characteristics of embedded systems for filtering high-volume sensor data before further processing. In particular, we investigate implementations of decision trees and random forests for the classical von-Neumann computing architecture and custom circuits by the means of field programmable gate arrays.

I. Introduction

Information technology is more and more integrated into every part of life, with applications ranging from factory monitoring, scientific experiments to serving consumer needs. Based on networking protocols and small, energy efficient, embedded systems it is now possible to measure and process data at virtually every place, at any time. In addition, combining multiple embedded systems creates one large ubiquitous computing system [1].