Active Diagnosis Through Information-Lookahead Planning

We consider challenging active diagnosis problems, that is, when smart exploration is needed to acquire information about a hidden target variable. Classical approaches rely on information-greedy strategies or ad-hoc algorithms for specific classes of problems. We propose to model this problem using the generic ρPOMDP formalism, which leads to an information-lookahead planning strategy, where the objective is to gather information-based reward. We empirically evaluate this approach on the Rock Diagnosis problem, which is a variation of the well-known Rock Sample problem, showing that we obtain better performance results than information-greedy techniques.