Explainability in Deep Reinforcement Learning, a Review into Current Methods and Applications

Tasks such as weather simulation, medical diagnosis, business optimisation and automation like autonomous cars have benefited from these new Artificial Intelligence (AI) methods. Some of these ML models are used in ways that their predictions can affect people's safety or commercial success. These models must be considered trustworthy with errors detected and dealt with before they can affect the success or safety of the process being controlled. Neural Networks (NNs), and in particular Deep Neural Networks (DNNs), represent one such class of ML algorithm. Due to the nature of DNNs, the decisions they produce can seem arbitrary. These DNNs are comprised of thousands of nodes that perform mathematical operations, creating a "black-box like" system, in which one is unable to judge the decisions being made by simple inspection.