Classifier Selection in Resource Limited Hardware: Decision Analysis and Resolution Approach

Abstract

Digitalization, Industry 4.0 and Internet of things (IoT) have become more popular in the recent years. Most of these systems depend on micro-controllers and sensors. These micro-controllers and sensors are mostly cheap, low RAM and low CPU systems; thus, they are resource constrained environments. In this study, a supervised learning classifier comparison technique suitable for resource constrained environments is proposed. This technique, Decision Analysis and Resolution (DAR), is originated in the domain of Software Engineering. First, DAR is explained using an example of car buying scenario. Then 11 off-the-shelf classifiers are compared using DAR for low RAM and less powerful CPU environments in an intrusion detection scenario. This scenario simulated on wellknown KDD99 intrusion detection dataset. All the experiments are realized using python scikit-learn package. According to the experiments, Decision Tree classifier is the most suitable to implement for resource constrained environments with a small lead. Results for the other three classifiers (Bagging, Multi Layer Perceptron, Random Forest) are also very similar. To aid the reproducibility of the experiments, the whole source code of the study is provided in the popular open source repository https://github.com/atiozgur/classifier-comparison-using-DAR.