Advisor

Christof Teuscher

Document Type

Report

Publication Date

8-19-2021

Subjects

Computational modeling, Multiagent systems, Radiation -- Measurement

Abstract

Advancement of radiation detection technology is an ongoing process, and adjustments are made based on pre-existing conditions of radiation presence--both natural and man made. Tools that are currently used for safely detecting radiation in urban environments exist in several forms: drones, robots, or handheld radiation detection devices. This is a harm reductive way to explore radiation-infected environments while preserving human health as best as possible. In order for these autonomous platforms to successfully detect radiation sources, an algorithm needs to be created that is capable of gathering crucial data on its own with little to no human interference. Machine learning has been the algorithm of choice for researchers, particularly reinforcement learning and deep reinforcement learning. These tools for information gathering are designed to have the algorithm “learn” on its own based on a reward system that allows for seekers called “agents” to fulfill its mission objective of radiation detection in urban environments. In this paper, we explore the capability of having multiple agents within a controlled environment learn to locate a radiation source on their own using a tactic called Differentiable Inter-Agent Learning. This concept would be built upon pre-existing work developed in a master’s thesis examining Proximal Policy Optimization for radiation source search using a single agent. By adding what could potentially be many agents to a radiation detection algorithm, it could provide a quicker and more efficient strategy for locating radiation sources remotely to preserve human health and safety. This multi-agent algorithm would examine if this is possible.

Persistent Identifier

https://archives.pdx.edu/ds/psu/36297

Share

COinS