AMHERST, Mass. – The Armstrong Fund for Science at the University of Massachusetts Amherst has announced its awards for 2018, which will grant $36,000 to an interdisciplinary research team for a two-year project to encourage transformative research on campus that introduces new ways of thinking about pressing scientific or technical challenges.
This year’s award goes to Shannon Roberts, assistant professor of mechanical and industrial engineering, and Philip Thomas, assistant professor of computer science, who have joined forces to launch a pilot study to address the question of when an automated driving system should warn a human driver that it may have to relinquish control of the vehicle in the near future. They will be recognized at the campus’s annual Awards Dinner on April 30.
As Roberts explains, if the system only gives warnings after it’s 100 percent certain that it will relinquish control, it may be too late for the human driver to respond. At the same time, if the warning is given when it is less than 100 percent certain, there may be so many false alarms that they degrade the driver’s trust in the system, making it less effective. “We seek a balance in the automated system between these two scenarios,” she says.
The researchers propose to use Thomas’s recently developed machine learning algorithms based on reinforcement learning to improve such systems. They also plan to use the driving simulator in the campus’s Human Performance Laboratory to test the effectiveness of their approach. Results will be used to submit grants to federal sources for further studies.
In particular, Thomas points out, “We study how the mechanism that determines when a warning will be given can be optimized using machine learning methods, so as to minimize the total number of resulting crashes.” He adds, “This is not merely the question of how many seconds are required before transferring control to a human.”
Rather, they point out, a combination of the current predicted time until control will be transferred to the human plus the current uncertainty is needed for the system to make an appropriate decision. There are many reinforcement learning algorithms, they add, with the ﬁrst created on campus in the 1980s, but most provide no practical safety guarantees, a critical limitation. “For high-risk applications like ours,” they say, they plan to test a special type of “safe reinforcement learning” designed by Thomas.
Vice Chancellor for Research and Engagement Michael Malone administers Armstrong grants in a competitive proposal process. Benefactors John and Elizabeth Armstrong established their Fund for Science in 2006 to identify and support promising research directions that do not yet have enough data available for the principals to apply to standard funding channels.