A team of U.S. Army researchers took another step to make the service’s multidomain operations concept a reality by making training reinforcement learning-based policies less unpredictable in order for ground robots and other autonomous systems to adapt to evolving conditions on the battlefield, the service reported Tuesday.
Alec Koppel, a researcher with DEVCOM Army Research Laboratory, and his team came up with policy search schemes for general utilities and found that those schemes were able to explore unknown domains in an efficient manner, integrate prior experience and lessen the volatility of reward accumulation.
With those schemes, Koppel said the research team took a step “towards breaking existing sample efficiency barriers of prevailing practice in reinforcement learning.” The Army conducted the research in partnership with Google Deepmind, Princeton University and University of Alberta.
“I am optimistic that reinforcement-learning equipped autonomous robots will be able to assist the warfighter in exploration, reconnaissance and risk assessment on the future battlefield,” Koppel said.
