Fluent learning: Elucidating the structure of episodes

Abstract

Fluents are logical descriptions of situations that persist, and composite fluents are statistically significant temporal relationships between fluents. This paper presents an algorithm for learning composite fluents incrementally from categorical time series data. The algorithm is tested with a large dataset of mobile robot episodes. It is given no knowledge of the episodic structure of the dataset (i.e., it learns without supervision) yet it discovers fluents that correspond well with episodes.