In many complex problems related to discovery, detection, and diagnosis, researchers and practitioners alike are continually faced with the question ?What data should I gather next?? When the possibilities for data collection are overwhelming, and experiments or measurements are costly or time-consuming, this question becomes all the more critical. This research investigates guided sensing algorithms, which make recommendations about the next measurements to gather, with the understanding that a domain expert makes the final decision. Motivated by applications in emergency response and high-throughput cell-based analysis, this work develops new methods for guided sensing that account for temporal and task-based constraints, missing data, and environmental noise as well as human error.
The research makes two primary technical contributions. First, it generalizes classical query-based learning algorithms to be robust to noise, with input and output designed to match users? needs. To accomplish this, greedy decision tree algorithms are designed with respect to new performance measures that reflect task-specific objectives and constraints. Second, this work develops interactive, nonparametric methods for statistical matching, a fundamental problem in data fusion. The approach is grounded in new methods for nonparametric clustering with missing data, and for unsupervised sequential experimental design.
