Co-Principal Investigators Shannon Roberts (Mechanical and Industrial Engineering Department) and Philip Thomas (College of Information & Computer Sciences) are receiving a $36,000 award from the 2018 Armstrong Fund for Science, administered by the UMass Amherst Office of Research Development. The grant will fund their two-year project entitled “Improving Warning Systems of Driving Automation Systems through Reinforcement Learning,” aimed at optimizing precisely when a driving automation system should safely alert a human driver to take control of the vehicle when approaching a hazard.
Roberts and Thomas will use the results from this pilot study to support a future proposal submitted to the National Science Foundation’s Computer & Information Science & Engineering Directorate.
The Armstrong Fund for Science was established in 2006 through a gift from John and Elizabeth Armstrong. Its goal is to encourage faculty at UMass Amherst to pursue research that has a significant likelihood of major science or engineering impact.
The crucial problem addressed by Roberts and Thomas in their Armstrong project is that driving automation systems sometimes encounter hazardous situations that they are not either equipped or programmed to handle properly.
“In these situations, the car transfers control back to a human driver,” as Roberts and Thomas explain. “However, it would be dangerous for a driving automation system to instantaneously transfer control to a human without any warning, as the human driver may not have an appropriate level of situation awareness with respect to the driving environment.”
The researchers reference studies that demonstrate how a driver must be alerted at least eight seconds before his or her vehicle approaches a hazard in order to ensure a safe amount of time for the human being to adjust to such a perilous situation. In this context, if the automated system waits until it is certain that it will have to transfer control, there might not be enough time for the human driver to respond appropriately. On the other hand, if the car warns the driver before it is certain it must transfer control, a “false positive warning” will often result in which the automated system ultimately determines not to hand over control to the human.
The two researches note that these false positive warnings can be just as dangerous as late warnings because they cause drivers to lose trust in the warning system, a situation which increases the chance that a driver will ignore or fail to respond to future warnings.
As Roberts and Thomas conclude, “Thus, the problem that we consider in this project is: When should a driving automation system begin to warn a human driver that s/he may need to resume manual control of the vehicle? More precisely, 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.”
Roberts and Thomas propose the use of a class of machine learning algorithms, called reinforcement learning algorithms, to optimize this mechanism within a driving automation system.
“Specifically, we will use algorithms recently developed at UMass,” explain Roberts and Thomas, “which provide high-probability guarantees that they will only increase performance; in this context, this corresponds to decreasing the number of crashes. After testing these methods completely via computer simulations, we will conduct a study with human participants in the Human Performance Lab's driving simulator to determine the efficacy of our approach.”
According to the website of the Armstrong Fund for Science, it “is intended for faculty members with aggressive research visions, who are willing to challenge conventions in their field. Grants from this program should be used to strategically expand one's research program by positioning it for large, extramural grants or key industrial partnerships. The research should represent a new initiative: either a bold, new line of research, or the application of prior research to a field that has no precedent for it.” (April 2018)