Learning elements of representations for redescribing robot experiences

This paper presents our first efforts toward learning simple logical representations from robot sensory data and thus toward a solution for the perceptual grounding problem [2]. The elements of representations learned by our method are states that correspond to stages during the robot’s experiences, and atomic propositions that describe the states. The states are found by an incremental hidden Markov model induction algorithm; the atomic propositions are immediate generalizations of the probability distributions that characterize the states. The state induction algorithm is guided by the minimum description length criterion: the time series of the robot’s sensor values for several experiences are redescribed in terms of states and atomic propositions and the model that yields the shortest description (of both model and time series) is selected.