Active Feature Acquisition with POMDP Models

We consider the problem of active feature acquisition (AFA), where the selection of a new feature is conditional on the instantiations of previously selected features. The problem is formulated as a partially observable Markov decision process (POMDP). We present a method to construct an approximate POMDP for the AFA problem and discuss its accuracy. We propose a non-stationary policy to improve the classiflcation performance at lower feature acquisition costs. The methods are demonstrated on medical diagnosis problems.