Comparison of Feature Selection Algorithms for Predicting Execution Time in Embedded Systems

Abstract

The advances in Machine Learning techniques promoted the development of AI modules integrated to control operations in embedded systems, thus, increasing the autonomous potential of such applications. In this paper, we present a comparison of feature selection algorithms over their ability to identify meaningful sub-sets of PMU counters to predict the execution time of tasks in embedded systems. The comparison is based on a data-set of performance events sampled in a real embedded platform during the execution of embedded systems benchmarks presenting representative embedded systems workloads. Our comparison features 10 different feature selection algorithms from both model-dependent
and model-independent classes. These algorithms are evaluated based on the accuracy achieved over 9 different Machine Learning methods, including models from the linear, non-linear, and decision tree classes trained with the events yielded by the feature selection process. The results indicate that the Successive Feature Selection feature selector presents the best overall performance, while the Conditional Mutual Information algorithm produced the best individual sub-set of features. The same algorithm also presented the second-best overall performance. In terms of the Machine Learning algorithms used for the evaluation, the XGBoost algorithm presented the best performance, followed by AdaBoost.