Applicability of feature selection on multivariate time series data for robotic discovery

Open ended robotic discovery aims at enabling robots to autonomously design and execute sophisticated experiments for gaining conceptual insight about real world. Such experiments are planned activities rather than innate motor commands and thus each single experiment results in a multivariate time series. In such a scenario, reducing the number of features in order to allow a symbolic learner to build a correct conceptual model of underlying phenomena is a fundamental task. Only few feature selection approaches deal with finding relevant features in multivariate time series, which is just what the robot receives through its sensors. In this paper, we present results of applicability of a range of feature selection and time series analysis approaches on a novel real world scenario for autonomous robotic discovery. We found that even sophisticated representations and state of the art techniques, which perform very well on other benchmarks, do not show significant results in context of open ended discovery.