On the Influence of Time-Correlation in Initial Training Data for Model-Based Policy Search

Model-based policy search techniques are a way to quickly learn real-world robotics behaviors. Indeed, learning a model allows the robot to imagine interactions with its environment and to furthermore refine its behavior without actually interacting with it. Nevertheless, a good model is critical for such approaches to succeed, and initial model training steps can change drastically the performance of the policy search algorithm. In this paper, we propose to study the impact of various initial data gathering methods for bootstrapping model training in a model-based policy search algorithm. We compare five initialization methods, two being used in the state of the art and the last three one representing various degrees of timecorrelation. We then show a link between the time-correlation of the initialization method and an environment metric that we call consistency, empirically demonstrating on three robotic systems that model prediction error can be predicted from environment consistency and initialization method action sequences timecorrelation. We finally show the impact the initialization method has on a state of the art model-based policy search algorithm, demonstrating consistent results with the proposed metric and the initialization methods prediction errors, allowing for a better choice of initial training data beforehand.