Degradation of Performance in Reinforcement Learning with State Measurement Uncertainty
Mark C. McKenzie, Mark D. McDonnell · 2019
We detail the use of open source training environments to investigate the applicability of standard reinforcement learning techniques to inherently error prone tasks expected in real world application of artificial intelligence. Numerical experiments were conducted in which the performance of both Q Learning and Policy Gradient agents' ability to obtain high reward was compared as the observation state measurement uncertainty was increased. The purpose of the research was to assess the applicability of reinforcement learning to real world applications of self-protection of military platforms, where it is expected that the observed state space is uncertain at best. We found in our experiments that Q Learning is more stable in the presence of state uncertainty than policy gradient learning.