Mutual Information Weighing for Probabilistic Movement Primitives

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