Mutual Information Weighing for Probabilistic Movement Primitives

Abstract

Reinforcement Learning (RL) of trajectory data has been used in several fields, and it is of relevance in robot motion learning, in which sampled trajectories are run and their outcome is evaluated with a reward value. The responsibility on the performance of a task can be associated to the trajectory as a whole, or distributed throughout its points (timesteps). In this work, we present a novel method for attributing the responsibility of the rewards to each timestep separately by using Mutual Information (MI) to bias the model fitting of a trajectory.