Regression Model Trees: Compact Energy Models for Complex IoT Devices

Abstract

The energy and timing behaviour of embedded components such as radio chips or sensors plays an important role when developing energy-efficient cyber-physical systems and IoT devices. However, datasheet values generally have low accuracy and may be incomplete, and performing new energy measurements after each code or hardware configuration change is time-consuming. While energy models – automatically generated from benchmarks exercising all relevant device configurations – offer a solution, they should have both low prediction error and low complexity in order to be useful to humans as well as energy simulations. With today’s increasingly complex devices and drivers, generating compact and accurate energy models is becoming harder due to non-linear effects and interdependencies between configuration parameters. To address this issue, we present Regression Model Trees. By combining software product line engineering and energy modeling methodologies, these are capable of automatically learning complex energy models from benchmark data. Using energy and timing benchmarks on two embedded radio chips and an air quality sensor, we show that Regression Model Trees are both more accurate than conventional energy models and less complex than state-of-the-art approaches from the product line engineering community. Thus, they are easier to understand and use for humans and algorithms alike. We observe two-to 100-fold complexity reduction, and a maximum energy model error of 6 % with cross-validation.