Reducing complexity in robotic learning by experimentation

In the learning by experimentation (LbE) paradigm, every knowledge is deduced from the analysis of experimental data, obtained with a proper experiment. The application of LbE in robotics is often limited by the excessive amount of sensor data. To face this problem a two-phases design strategy is possible. The former phase, called feature selection, isolates among the quantities involved in the learning; while the latter phase designs an experiment considering only a reduced subset of variables. This paper proposes a feature selection method: the feature section problem is formulated as a conditional independence problem and it is handled by applying the contingency table theory. The goodness of the work is tested on a LbE-based framework, extended to fit the presented method, regardless of the nature of the knowledge.