Correlating Sensors and Activities in an Intelligent Environment: A Logistic Regression Approach Fahd Albinali 1 (albinali@cs.arizona.edu) Prasad Boddupalli 1 (bprasad@cs.arizona.edu) Nigel Davies 1, 2 (nigel@comp.lancs.ac.uk) Adrian Friday 2 (adrian@comp.lancs.ac.uk) 1 Dept. of Computer Science University of Arizona, Arizona, USA 2 Computing Department Lancaster University Lancaster, UK ABSTRACT An important problem in intelligent environments is how the system can identify and model users’ activities. This paper describes a new technique for identifying correlations between sensors and activities in an intelligent environment. Intelligent systems can then use these correlations to recognize the activities in a space. The proposed approach is motivated by the need for distinguishing the critical set of sensors that identifies a specific activity from others that do not. We compare several correlation techniques and show that logistic regression is a suitable solution. Finally, we describe our approach and report preliminary results. Keywords Ambient intelligence, activities, correlation, regression INTRODUCTION In his classic paper “The Computer for the 21 st Century” [14] Weiser envisions a world of intelligent environments that are highly aware of their inhabitants. In this vision, physical spaces are enhanced with computing capabilities to act more intelligently: they observe, interact with and react to humans in meaningful ways. They understand human reasoning, analyze behaviors and infer intentions. Furthermore, intelligent environments actively collaborate with their inhabitants to assist them in making their surroundings more pleasant. Intelligent environments even take decisions and execute actions on their own. They become integral participants in the daily human activity. A critical element that Weiser anticipated, yet has not been achieved, is the invisibility of pervasive systems. The ability of such systems to disappear into the background of everyday life is dependant on their ability to correctly interpret the state of the environment and to act accordingly: intelligent systems that incorrectly interpret the state of the world or the intentions of users are likely to take inappropriate actions that are not naturally anticipated by users [6]. Such incorrect actions could become very disruptive and intrusive to users, they distract the inhabitants of intelligent spaces from their ongoing activity and therefore, they make them more aware of the system. This paper begins to address the challenge of designing less intrusive intelligent environments that can engage in richer and more meaningful interactions with users. We believe that such systems must have a deep understanding of user context and, specifically, should have an understanding of activities that a user is engaged in. Our approach is thus inspired by concepts from activity theory [9] and requires support for three basic system functions: • Sensing context: By observing and monitoring users’ context, intelligent systems can collect information about the intelligent space and its inhabitants. • Analyzing context: By analyzing users’ context, intelligent systems can estimate and interpret users’ activities. • Gracefully reacting to the inhabitants: By understanding users’ activities, intelligent systems can react unobtrusively to their inhabitants and therefore can potentially become more invisible. In this paper, we focus on one aspect of our system design, i.e. how to identify sensors that correlate with activities in an intelligent space. First, we motivate our use of an activity-centric approach and justify the need for precisely identifying correlations between sensors and activities. Second, we identify a number of desirable properties for activity-aware intelligent systems. We then analyze different techniques for identifying the correlations between sensors and activities and show that statistical logistic regression