Learning Relational Grammars from Sequences of Actions

摘要

Many tasks can be described by sequences of actions that normally exhibit some form of structure and that can be represented by a grammar. This paper introduces FOSeq, an algorithm that learns grammars from sequences of actions. The sequences are given as low-level traces of readings from sensors that are transformed into a relational representation. Given a transformed sequence, FOSeq identifies frequent sub-sequences of n-items, or n-grams, to generate new grammar rules until no more frequent n-grams can be found. From m sequences of the same task, FOSeq generates m grammars and performs a generalization process over the best grammar to cover most of the sequences. The grammars induced by FOSeq can be used to perform a particular task and to classify new sequences. FOSeq was tested on robot navigation tasks and on gesture recognition with competitive performance against other approaches based on Hidden Markov Models.