default Deep learning versus expert reasoning under uncertainty

Deep learning versus expert reasoning under uncertainty. Topple, Jessica M.; Ben Abdallah, Nadia; Jousselme, Anne-Laure. CMRE-MR-2019-002. April 2020.

This report presents a deep learning approach to a situational awareness task using a recurrent neural network implementation. The network was developed for The Risk Game, a methodology for capturing expert knowledge and reasoning under uncertainty while performing a maritime situational awareness task. The game provides information from multiple sources with variable information quality, abstracted by cards to the human expert players and by feature vectors to the deep learning network. Belief in each of two possible vessel identities is assessed by both the humans and the network given the sequence of information cards or feature vectors provided. This report compares the vessel identity belief assessments of the experts to the network predictions and demonstrate that such deep learning networks hold great potential as general situation awareness and decision making support systems, and as baseline automatic reasoners from which to compare and model human reasoning behaviour.