Application of Computational Intelligence for Command & Control of Unmanned Air Systems
2019 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved. Computational Intelligence is a field of study devoted the creation of algorithms that can be implemented in a software agent to accomplish complex, specified tasks. There are many machine learning algorithms, each with their own advantages and disadvantages. For example, relatively simple Q-Learning algorithms have been shown to be successful in controlling Unmanned Air System (UAS) motion video tracking of ground targets. However, updating these learning algorithms onboard a UAS poses a number of inherent challenges with both hardware and software. This paper investigates the various challenges to online machine learning onboard UAS and considers proposed solutions with simulation proof of concept. Methods to address these challenges will be explored by developing a change from static, offline trained Q-learning algorithms to dynamic, online learning algorithms of the reinforcement learning class. Application of dynamic learning algorithms significantly changes the implementation. A planar motion aircraft simulation was built to train and test reinforcement learning algorithms for tracking ground vehicles. The improved training approach will be demonstrated using a model trained with new online reinforcement learning techniques. This will be compared qualitatively to a basic Q-learning model created from the same environment. These changes, and their implications with regard to the main challenges mentioned initially, will be presented along with proposed extensions to this work.