Towards Self-Improving Activity Recognition Systems Based on Probabilistic, Generative Models
Martin Jänicke, Sven Tomforde, Bernhard Sick · 2016
Activity Recognition (AR) systems increasingly pervade our daily lives, reaching from the monitoring of daily activities to the support in medical care. However, such systems are created with narrow system specifications and a very specific field of application in mind, leading to application-dependent setup and configuration by their users. To overcome these limitations, research focuses on autonomous solutions that are able to work with no (or minimal) user interaction. A major step towards this goal is the integration of novel input sources (i.e., sensors) at runtime, leading to a dynamic input-space (i.e., variable dimensionality). This paper presents an approach to systematically investigate methods necessary for the creation of self-adapting classification systems. This includes an architecture, based on Organic Computing principles, as well as the development of measures for comparing probabilistic models and procedures for evaluating classifiers of different dimensionality. With such evaluation techniques, systems are enabled to adapt their system model at runtime in a self-organized manner. Besides self-improvement (adding a new sensor), we also address the problem of self-healing (replacing a sensor that dropped out).