Bipartite Graph Learning for Autonomous Task-to-Sensor Optimization
Abstract
This study explores the question of how machine learning can be applied to identify the most appropriate sensor for a task by optimizing task-to-sensor matching. The concept of a bipartite graph provides a mathematical framework for task-to-sensor mapping by establishing connectivity between various high-level tasks and the specific sensors and/or processes that must be invoked to fulfil those tasks and other mission requirements. The connectivity map embedded in the bipartite graph can change depending on the availability/unavailability of resources, the presence of constraints (physics, operational, sequencing), and the satisfaction of individual tasks. All of these considerations may be encoded in the value matrix of the graph. Changes can also occur according to the valuation, re-assignment and re-valuation of the perceived task benefit and how the completion of a specific task (or group of tasks) can contribute to the state of knowledge prompting the periodically re-solve the matching. The results of this study show that a deep neural network architecture can be used to solve the bipartite matching problem in an autonomous fashion. The scalability of the problem is demonstrated using an 800 by 800 graph. Learning acceleration is also studied and it is recommended that this aspect be further explored as part of future investigations.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2022
- Accession Number
- AD1184537
Entities
People
- Mark Karpenko
- Michael Zepeda
- Ronald J. Proulx
Organizations
- Naval Postgraduate School