Fluents are logical descriptions of situations that persist, andcompositefluents are statistically significant temporal relationships(nearlyidentical withthosein Allen’stemporal calculus) betweenfluents. Thispaperpresentsan algorithm for learningcompositefluents incrementally.Thealgorithm is tested witha large datasetof mobilerobotepisodes.The algorithmis givennoknowledgeof the episodicstructure of the dataset (i.e., it learns withoutsupervision) yet discoversfluents that correspondwell withepisodes.More generally,the algorithmelucidateshiddenstructure in time seriesof binaryvectors.
Paul R. Cohen