Quantifying the effect of feedback frequency in interactive reinforcement learning for robotic tasks

Reinforcement learning (RL) has become widely adopted in robot control. Despite many successes, one major persisting problem can be very low data efficiency. One solution is interactive feedback, which been shown to speed up RL considerably. As a result, there an abundance of different strategies, are, however, primarily tested on discrete grid-world and small scale optimal control scenarios. In the literature, no consensus about feedback frequency or at time most beneficial. To resolve these discrepancies we isolate quantify effect robotic tasks with continuous state action spaces. The experiments encompass inverse kinematics for manipulator arms complexity. We show that seemingly contradictory reported phenomena occur complexity levels. Furthermore, our results suggest single ideal exists. Rather should changed as agent's proficiency task increases.