Deep Influence Diagrams: An Interpretable and Robust Decision Support System

摘要

Interpretable decision making frameworks allow us to easily endow agents with specific goals, risk tolerances, and understanding. Existing decision making systems either forgo interpretability, or pay for it with severely reduced efficiency and large memory requirements. In this paper, we outline DeepID, a neural network approximation of Influence Diagrams, that avoids both pitfalls. We demonstrate how the framework allows for the introduction of robustness in a very transparent and interpretable manner, without increasing the complexity class of the decision problem.