Bayesian Clustering of Sensory Inputs by Dynamics

This paper describes a Bayesian approach to the abstraction of sensor dynamics using a new clustering algorithm for time series to learn prototypical behaviors of a robot's sensory inputs. Each sensor stream reading is modeled as a Markov chain (mc). The abstraction process is performed by an unsupervised clustering algorithm returning the most probable set of clusters capturing the robot's sensory experiences. In order to increase e ciency, the algorithm uses an heuristic search strategy merging the closest mcs according to a measure of similarity based on entropy.