When Fei Fang was a graduate student she was introduced to game theory—mathematical models that describe strategic interactions among rational decision-makers. The IEEE member knew she had found her calling. She has combined the modelling technique with machine learning to thwart terrorist attacks and reduce animal poaching.
For Fang’s work in the field, she was named one of IEEE Intelligent Systems magazine’s “AI’s 10 to Watch in 2020.”
Fang, an assistant professor of computer science at Carnegie Mellon, is now working with 412 Food Rescue, a nonprofit in Pittsburgh, to improve its system for alerting volunteers when surplus food is available for pickup.
“Food waste is a huge problem, and there are a lot of people suffering from food insecurity in Pittsburgh,” Fang says. The nonprofit “is such a cool organization, doing a really amazing thing, and I thought that maybe I could help.”
ATTENTION TO SAFETY
Fang started programming as a student at Changzhou Senior High School, in China. She says it intrigued her because “you can do something that can have a tangible impact.”
Although she enjoyed coding, she decided to pursue a degree in biology at Tsinghua University, in Beijing. She eventually changed her major to electrical engineering because, she says, she found it more interesting. She graduated in 2011 with a bachelor’s degree, then went to the United States to pursue a doctorate at the University of Southern California, in Los Angeles.
There, Fang joined computer science professor Milind Tambe’s team and led a research project that sought to use computational game theory to help the U.S. Coast Guard plan patrol routes to protect the Staten Island Ferry system from terrorist attacks.
A similar AI algorithm had been used for airport security since 2007 to schedule canine patrols. She and her research colleagues planned on applying the system to the ferry service but found they could not. For the airport algorithm to work for seaport security, she says, it had to have spatial temporal reasoning: the ability to take into account the movement of the ferries and patrol boats.
“We had to deal with an infinite number of patrol routes the U.S. Coast Guard patrol boats could take,” she says. “The attacker could also strike the ferries at any time. This makes the problem very different [from airport security] and more challenging because there are an infinite number of actions possible on both sides.”
She and her team came up with a compact representation in which the AI algorithm considers all the times and locations terrorists might attack. The Coast Guard deployed the PROTECT (Port Resilience Operational/Tactical Enforcement to Combat Terrorism) model in 2013, and it is still in use.
After the successful deployment of the ferry safety algorithm, Fang explored other problems AI could solve. Under Tambe’s leadership, she and her colleagues developed a system to stop animal poachers on wildlife preserves before they strike.
The machine-learning system, dubbed PAWS (Protection Assistant for Wildlife Security), uses data from past patrols to predict where poaching is likely to occur and a game-theory model to help generate randomized, unpredictable patrol routes, according to a 2018 IEEE Spectrum article on the project.
Save the Wildlife, Save the Planet: Protection Assistant for Wildlife Security (PAWS)www.youtube.com
The first PAWS trial was conducted in Malaysia in 2014. Fang and her team sent rangers to protected areas where endangered tigers live. The rangers found footprints and other signs of human activity, Fang says. But the trial also showed that rangers couldn’t follow the routes in a straight line, like the system recommended, because of the terrain. Fang and her colleagues refined their system and began testing the improved tool, which took the topography into account when generating recommended patrol routes.
Fang earned her Ph.D. in computer science in 2016 and joined Harvard as a postdoctoral fellow. But she continued to work with Tambe on PAWS.
More trials were conducted in Uganda in 2016 in collaboration with the Queen Elizabeth National Park’s Wildlife Conservation Society and in China in 2017 and 2019 with the World Wildlife Foundation. Park rangers were sent to locations that PAWS predicted to be poaching hotspots that were not frequently patrolled. The rangers found several snares used to catch animals, Fang says.
Fang shows off a tiger snare discovered in a wildlife preserve in northeast China.Yongchao Jing
She participated in a two-day field test of the system in China, and the experience inspired her to add a capability to PAWS. The new feature helps rangers make decisions while on patrol. If rangers find poachers’ footprints during a patrol, for example, they can use the system to decide whether they should deviate from their original route and follow the trail.
This year Microsoft added PAWS to its Azure platform, a portfolio of AI services designed for developers and data scientists. It also was integrated into SMART, a platform that consists of software and analysis tools designed to help conservationists manage and protect wildlife. SMART is used at more than 600 conservation sites around the globe.
“I’m really happy to see that part of our algorithm has been integrated into the software and is now available worldwide,” Fang says.
Fang left Harvard in 2017 to join Carnegie Mellon as an assistant professor.
Food waste and food insecurity are huge problems around the world, including Pittsburgh. Fang and her research team at Carnegie Mellon’s Institute for Software Research are developing an algorithm to help 412 Food Rescue increase the pickup rate of good but unsellable food from grocery stores. Volunteers pick up the food and deliver it to homeless shelters and people in need.
The nonprofit uses a smartphone app to notify volunteers who are within 8 kilometers of the pickup zone when there is food available. But, according to Fang, it is not guaranteed that a volunteer will accept the task. The organization’s co-founder and chief executive, Leah Lizarondo, turned to Fang for help to make the process more efficient.
Volunteers were notified through the app’s push notification. If no one accepted the request after 15 minutes, a volunteer would text or call other volunteers to find out if they could pick it up.
“The first idea we had was to try to use machine learning to predict which food rescue request might be at risk of not being accepted by volunteers,” Fang says. “Once those orders were identified, the dispatchers in the organization could take action and try to contact volunteers they know, or see if there are other ways that the food could be delivered.”
The second idea, devised by her student Ryan Shi, was to improve the push notification scheme. He realized that the 8-km radius and 15-minute wait time might not be the best options, so the team developed an algorithm to determine the optimal choices based on historical data. Now the scheme’s radius is 8.8 km and the waiting time is 16 and a half minutes.
The new system was deployed in February 2020. In one month, the pickup rate increased from 84 percent of food being claimed to 88 percent. The time it took for a task to be accepted decreased from 78 minutes to 43 minutes.
Fang and her team are further improving the system by making it possible for the algorithm to suggest specific volunteers who are more likely to accept and pick up food from certain stores.
Fang joined IEEE this year so that she could keep current with technological advances through the organization’s publications, including IEEE Spectrum.
“I also wanted the opportunity to publish my work in IEEE journals,” she says. You can find several of her articles in the IEEE Xplore Digital Library.
She says she hopes to collaborate with fellow IEEE members on projects.
Joanna Goodrich is the associate editor of The Institute, covering the work and accomplishments of IEEE members and IEEE and technology-related events. She has a master's degree in health communications from Rutgers University, in New Brunswick, N.J.