One problem of fundamental interest to the artificial intelligence and machine learning research communities is the modeling of the state of the everyday human environment. Reasoning in a "commonsense" manner about everyday life based on sensory input presents some significant challenges. First, there is the problem of effectively collecting and representing commonsense information about the state of the world, including the objects, concepts, actions, and state variables that may be necessary to fully model the environment and the relationships between these entities. An effective model of the environment may require a very large amount of such data, which may be difficult to collect, to organize, and to evaluate. Second, there is the problem of effectively reasoning over this information once collected. Performing inference about the state of the environment given evidence may require a large amount of computation over a great many variables.

In this work, we present a system, SRCS, which is designed to model and reason about the state of the everyday human environment using large scale commonsense reasoning. SRCS makes use of lightweight wearable sensors to observe human activity, and incorporates statistical learning techniques and preexisting sources of commonsense data to produce a graphical model of the state of the human environment over time. To improve the effectiveness of our system, we use semi-supervised machine learning techniques to learn a model based on sparsely labeled training data. Finally, to improve the efficiency of this model, we introduce a means of inferring a context - i.e. a subset of particularly relevant information - to the observations the system may encounter, and use this context discovery to produce more efficient reasoning. We present experimental results that confirm the efficacy of our techniques, and conclude with current directions of our research and future work.