UXV NETWORKED CONTROL SYSTEM OPTIMIZATION AND EXPLAINABILITY VIA SUPERVISED LEARNING

Abstract

This thesis explores the application of supervised machine learning techniques into improving the performance and interpretability of networked control systems (NCS). Unmanned Systems (UxVs) can offer quick and efficient target location through extended reconnaissance missions covering large areas. Controlling a group of UxVs in a collaborative manner is a difficult problem, and one solution in development at the Naval Postgraduate School is Behavior Integrated Optimization for Networked control systems (BION). BION is an NCS that leverages a Partially Observable Monte Carlo Planning (POMCP) algorithm to provide near-optimal near-real-time formation control. This thesis aims to resolve certain limitations of POMCP algorithms by approximating key functionality with a trained neural network. The neural network leverages probabilistic techniques and saliency maps to improve interpretability, while also bringing improvements to processing speed. Experiments were performed on different architectures based on ResNet, and the trained neural network was successfully integrated into BION. This research expands our ability to experiment and develop BION, which furthers the development and integration of UxVs in the fleet.