Premium
Evaluating trade‐offs in energy‐efficient error detection
Author(s) -
Rokicki Markus,
Drozda Martin
Publication year - 2015
Publication title -
international journal of communication systems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.344
H-Index - 49
eISSN - 1099-1131
pISSN - 1074-5351
DOI - 10.1002/dac.3028
Subject(s) - computer science , classifier (uml) , decision tree , false positive paradox , artificial intelligence , machine learning
Summary Due to the fact that in well‐engineered systems, normal behavior is the dominant case, it is necessary that normal behavior can be recognized in an energy‐efficient way. When cascading classification is applied, the task of the first classifier is to impose energy efficiency, whereas the task of the other classifier (or classifiers) is to provide false positives control. Adaptivity to novel or re‐occurring errors can be in this setup carried out by including a feedback loop enabling that the first classifier improves its capability to recognize normal behavior. However, in order to sustain adaptivity, it is necessary to provide the feedback loop with some minimum number of instances from both behavior classes. This results in a trade‐off between energy efficiency and error detection performance. In order to obtain a trade‐off that is suitable in a given situation, it is necessary that a methodology for comparing trade‐offs across qualitatively different types of error detection approaches is available. Therefore, we propose and evaluate an approach for comparing trade‐offs that is based on iso‐performance lines. In addition, we also propose and evaluate several modifications to k ‐nearest neighbor and decision tree classifiers, so that different grades of trade‐offs can be obtained. Copyright © 2015 John Wiley & Sons, Ltd.