Title:Group-based Query Learning for rapid diagnosis in time-critical situations

Abstract: In query learning, the goal is to identify an unknown object while minimizing the number of "yes or no" questions (queries) posed about that object. We consider three extensions of this fundamental problem that are motivated by practical considerations in real-world, time-critical identification tasks such as emergency response. First, we consider the problem where the objects are partitioned into groups, and the goal is to identify only the group to which the object belongs. Second, we address the situation where the queries are partitioned into groups, and an algorithm may suggest a group of queries to a human user, who then selects the actual query. Third, we consider the problem of query learning in the presence of persistent query noise, and relate it to group identification. To address these problems we show that a standard algorithm for query learning, known as the splitting algorithm or generalized binary search, may be viewed as a generalization of Shannon-Fano coding. We then extend this result to the group-based settings, leading to new algorithms. The performance of our algorithms is demonstrated on simulated data and on a database used by first responders for toxic chemical identification.