In comparison-based search task we are interested in building an active learning system that would assist a user in finding a specific item t in a database of objects by presenting him a pair of objects (i,j) and asking "is object i more similar to your target t than object j?". After we receive his answer, a new pair of objects is presented to the user and so on until we locate the desired target item t. How should we choose which pair of objects to show next in order to ask as few questions as possible?
All previous approaches were using information-theoretic active learning to find the pair of objects (i,j) to query next. In this project we would like to explore the potential of using reinforcement learning techniques in order to find which items to pick for the next question. The main challenges to be tackled are (i) the environment definition and (ii) the large action space (N^2, where N is the number of items).
To apply, please send your transcript and CV.