Observation Decoding with Sensor Models: Recognition Tasks via Classical Planning

Abstract

[EN] Observation decoding aims at discovering the underlyingstate trajectory of an acting agent from a sequence of observa-tions. This task is at the core of various recognition activitiesthat exploit planning as resolution method but there is a gen-eral lack of formal approaches that reason about the partialinformation received by the observer or leverage the distri-bution of the observations emitted by the sensors. In this pa-per, we formalize the observation decoding task exploiting aprobabilistic sensor model to build more accurate hypothesisabout the behaviour of the acting agent. Our proposal extendsthe expressiveness of former recognition approaches by ac-cepting observation sequences where one observation of thesequence can represent the reading of more than one variable,thus enabling observations over actions and partially observ-able states simultaneously. We formulate the probability dis-tribution of the observations perceived when the agent per-forms an action or visits a state as a classical cost planningtask that is solved with an optimal planner. The experimentswill show that exploiting a sensor model increases the accu-racy of predicting the agent behaviour in four different con-texts