Energy-aware reasoning agents for the mobile semantic web

Abstract

Description logics serve as a core component of the Semantic Web and are a means of representing knowledge to machines in a structured way so that software called reasoners can make logical inferences in a decidable fashion. In the last decade, a number of semantic technology deployments have been explored on mobile devices, yet little attention has been paid to understanding the energy and power requirements of these systems.

Based on the success of the prediction models, we develop a framework for building mobile energy-aware reasoners that, prior to the start of a realization task, can choose to offload computation to a remote service and show that this provides significant time, energy, and power savings. To simplify access to the results of this investigation, we present an extensible metric ontology built on best-in-class ontologies to encode ontology metrics as structured metadata and a web service that, given an ontology, generates time, energy, and power usage to support future energy-aware mobile semantic agents. We conclude with discussion of the implications of our findings for mobile semantic technologies and recommend practices for semantic web practitioners looking to deploy solutions using mobile technologies.

To address this lack of information, we present a hardware-based benchmark for evaluating description logic reasoners on mobile devices for energy and power consumption. We evaluate the performance of instance realization, the task of determining the classes to which an instance belongs, of 10 reasoner configurations on a broad range of description logic knowledge bases. The experimental data demonstrate that there are significant differences in behavior between reasoners and the task of choosing a reasoner cannot be reduced to simply picking a single ‘best’ reasoner. To exploit these data, we train a series of classification and regression models over a suite of ontology metrics found in the literature along with six new realization-oriented metrics. These predictive models demonstrate the ability to choose reasoners such that they improve 23.6% relative to an oracle that perfectly predicts energy use; this is equivalent to a gain of 1.9 hours of battery life over 8 hours of continuous use. We further demonstrate empirically that the “no free lunch” theorem for optimization problems also applies to the task of parallelizing description logic inferencing, in the sense that while parallelizing inference can improve query answer time, it comes at the cost of higher energy and power consumption.