Relational Sequential Inference with Reliable Observations

Abstract

We present a trainable sequential-inference technique for processes with large state and observation spaces and relational structure. Our method assumes ``reliable observations'', i.e. that each process statepersists long enough to be reliably inferred from the observations itgenerates. We introduce the idea of a ``state-inference function'' (fromobservation sequences to underlying hidden states) for representing knowledgeabout a process and develop an efficient sequential-inference algorithm, utilizing this function, that is correct for processes that generate reliableobservations consistent with the state-inference function. We describe arepresentation for state-inference functions in relational domains and give acorresponding supervised learning algorithm. Experiments, in relational video interpretation, show that our technique provides significantly improved accuracy and speed relative to a variety of recent, hand-coded, non-trainable systems.