Distributed Learning for Autonomous Vehicles in Communication-Constrained Environments

Nicole Cruz, Jose Fuentes, Leonardo Bobadilla

Abstract: Intelligent Autonomous Systems (IAS), which encompass Unmanned Aerial Systems (UAS), Unmanned Surface Vehicles (USV), and Unmanned Underwater Vehicles (UUV), are set to play a pivotal role across multiple domains. However, due to the demanding nature of their missions (e.g., disaster response and underwater exploration), common communication methods are often limited. This makes it essential to use efficient communication strategies for learning and decision-making. This paper introduces a novel approach, leveraging distributed machine learning algorithms along with data projection to achieve sublinear communication costs. Particularly, we incorporate Convolutional Neural Networks for feature extraction prior to projection, significantly enhancing efficiency and accuracy in autonomy tasks within bandwidth-constrained environments. This method uses the Johnson-Lindenstrauss transform for dimensionality reduction of both data and model parameters, facilitating efficient inter-agent communication by transmitting only projected model parameters. We further outline theoretical underpinnings showcasing the Johnson-Lindenstrauss transform’s effectiveness in maintaining convergence properties for Online Gradient Descent algorithms. Through experimental validation using image recognition tasks relevant to maritime surveillance, our approach demonstrates substantial reductions in communication cost while preserving high model performance.