Decision-theoretic active sensing for autonomous agents

Classification is a sub-task common to many problems faced by autonomous agents. Traditional treatment of classification in the Machine Learning literature assumes that a feature vector is given as input. This ignores the essential role of an autonomous agent as a proactive information gatherer. In this paper, we present a framework for making optimal sensing and information gathering decisions with respect to classification goals by formulating the problem as a partially observable Markov decision process and solving for the optimal policy. We demonstrate the utility of this approach on a simulated meteorite collection task faced by an autonomous rover.