CSC News

Creating a Virtual World-Building Machine for Military Training

Matt Shipman | News Services | 919.515.6386

Dr. James Lester | 919.515.7534

The Abstract Blog

The U.S. Army uses virtual simulation environments to train its soldiers in how to respond to a wide variety of situations, but building each of those scenarios from scratch is both expensive and time consuming. However, customized virtual worlds to address any training scenario may be right around the corner.

A cooperative agreement between the Army Research Laboratory and computer science researchers at NC State University aims to develop a program for generating customizable virtual training scenarios. They call the program “DeepGen.”

“Our goal is to improve the quality and reduce the cost of these training tools, in order to help soldiers develop the skills they need to stay alive and accomplish their objectives on the battlefield,” says James Lester, a distinguished professor of computer science at NC State and principal investigator on the DeepGen project.

“In addition to creating scenarios, we want to make DeepGen dynamic,” says Bradford Mott, a senior research scientist at NC State involved with the project. “We want DeepGen to create scenarios that not only incorporate targeted learning objectives, but reflect the strengths and weaknesses of individual trainees, based on each trainee’s demonstrated competencies up to that point.”

The scenarios created by DeepGen will be run on the existing Virtual Battlespace 3 platform, which was developed by Bohemia Interactive Simulations and is already in use by the Army.

Initially, the researchers will focus on having DeepGen generate scenarios to train Army personnel on how to call in artillery fire.

To that end, the researchers are collaborating with a company called Intelligent Automation Inc. to determine which variables need to be addressed in the relevant scenarios – weather, visibility, physical environment and so on.

“The DeepGen system will also continue to grow and change over time, using a machine learning technique called deep reinforcement learning to improve the scenarios it creates based on how trainees interact with each scenario,” says Jonathan Rowe, a research scientist on the NC State team.

“We’ll begin the process by running groups of virtual trainees through scenarios produced by the system, then through a group of what will essentially be beta testers,” Rowe says. “But the process will continue even after it’s in day-to-day use as a training tool. It will only get better.”

~shipman~