USING FISHER INFORMATION TO CREATE SELF-REFLECTION IN AUTONOMOUS SYSTEMS

Abstract

In modern society, the number and popularity of autonomous systems are increasing, and it seems certain that their importance will grow in the future. As early as 2017, Amazon was already working with more than 100,000 warehouse robots, and many companies have begun shipping with drones or autonomous vehicles around the world. In the future, autonomous systems are likely to play a major role not only in the public sector but also in the defense sector. In fact, the Republic of Korea Army introduced a "drone-bot" force in 2018, for defense applications. Nevertheless, the operation of autonomous systems poses several challenges. One is deciding how the autonomous system will make decisions in an uncertain situation. What if the collected data is scarce, contains extreme values, and follows an unknown distribution? In light of these uncertainties, a robust estimation method is needed. Autonomous systems should make judgments that lead to decisions that not only yield the good results but also, more importantly, avoid catastrophic outcomes. In this thesis, we present two fast and conservative estimation methods using Fisher information that adapt to the quality and quantity of the data. We compare our two methods with parametric estimates and maximum likelihood estimation under normal, log-normal, and exponential distributions. Finally, we apply the two methods to predict whether an unmanned underwater vehicle can successfully perform a mission.