Estimating Risk and Uncertainty in Deep Reinforcement Learning [article]

William R. Clements, Bastien Van Delft, Benoît-Marie Robaglia, Reda Bahi Slaoui, Sébastien Toth
2020 arXiv   pre-print
Reinforcement learning agents are faced with two types of uncertainty. Epistemic uncertainty stems from limited data and is useful for exploration, whereas aleatoric uncertainty arises from stochastic environments and must be accounted for in risk-sensitive applications. We highlight the challenges involved in simultaneously estimating both of them, and propose a framework for disentangling and estimating these uncertainties on learned Q-values. We derive unbiased estimators of these
more » ... es and introduce an uncertainty-aware DQN algorithm, which we show exhibits safe learning behavior and outperforms other DQN variants on the MinAtar testbed.
arXiv:1905.09638v5 fatcat:iq4savzrgfdqlovzzkowgfkdbe