Learning patterns of states in time series by genetic programming

A state in time series can be referred as a certain signal pattern occurring consistently for a long period of time. Learning such a pattern can be useful in automatic identification of the time series state for tasks like activity recognition. In this study we showcase the capability of our GP-based time series analysis method on learning different types of states from multi-channel stream input. This evolutionary learning method can handle relatively complex scenarios using only raw inputs requiring no features. The method performed very well on both artificial time series and real world human activity data. It can be competitive comparing with classical learning methods on features