CMU Scientists Solve 90-Year-Old Geometry Problem

Math Puzzle Resolved by Translating It Into Satisfiability Problem

Carnegie Mellon University computer scientists and mathematicians have resolved the last, stubborn piece of Keller's conjecture, a geometry problem that scientists have puzzled over for 90 years. By structuring the puzzle as what computer scientists call a satisfiability problem, the researchers put the problem to rest with four months of frenzied computer programming and just 30 minutes of computation using a cluster of computers. "I was really happy when we solved it, but then I was a little sad that the problem was gone," said John Mackey, a teaching professor in the Computer Science Department (CSD) and Department of Mathematical Sciences who had pursued Keller's conjecture since he was a grad student 30 years ago. "But then I felt happy again. There's just this feeling of satisfaction." The solution was yet another success for an approach pioneered by Marijn Heule, an associate professor of computer science who joined CSD last August. Heule has used an SAT solver — a computer program that uses propositional logic to solve satisifiability (SAT) problems — to conquer several hoary math challenges, including the Pythagorean triples problem and Schur number 5. "The problem has intrigued many people for decades, almost a century," Heule said of Keller's conjecture. "This is really a showcase for what can be done now that was not possible previously." The conjecture, posed by German mathematician Eduard Ott-Heinrich Keller, has to do with tiling — specifically, how to cover an area with equal-size tiles without any gaps or overlap. The conjecture is that at least two of the tiles will have to share an edge and that this is true for spaces of every dimension. It's easy to prove that it's true for two-dimensional tiles and three-dimensional cubes. As of 1940, the conjecture had been proven true for all dimensions up to six. In 1990, however, mathematicians proved that it doesn't work at dimension 10 or above. That's when Keller's conjecture captured the imagination of Mackey, then a student at the University of Hawaii. With an office next to the university's computing cluster, he was intrigued because the problem could be translated, using discrete graph theory, into a form that computers could explore. In this form, called a Keller graph, researchers could search for "cliques" — subsets of elements that connect without sharing a face, thus disproving the conjecture. In 2002, Mackey did just that, discovering a clique in dimension eight. By doing so, he proved that the conjecture fails at that dimension and, by extension, in dimension nine. That left the conjecture unresolved for dimension seven. When Heule arrived at CMU from the University of Texas last year, he already had a reputation for using the SAT solver to settle long-standing open math problems. "I thought to myself, maybe we can use his technique," Mackey recalled. Before long, he began discussing how to use the SAT solver on Keller's conjecture with Heule and Joshua Brakensiek, a double major in mathematical sciences and computer science who is now pursuing a Ph.D. in computer science at Stanford University. An SAT solver requires structuring the problem using a propositional formula — (A or not B) and (B or C), etc. — so the solver can examine all of the possible variables for combinations that will satisfy all of the conditions. "There are many ways to make these translations, and the quality of the translation typically makes or breaks your ability to solve the problem," Heule said. With 15 years of experience, Heule is adept at performing these translations. One of his research goals is to develop automated reasoning so this translation can be done automatically, allowing more people to use these tools on their problems. Even with a high quality translation, the number of combinations to be checked in dimension seven was mind-boggling — a number with 324 digits — with a solution nowhere in sight even with a supercomputer. But Heule and the others applied a number of tricks to reduce the size of the problem. For instance, if one data configuration proved unworkable, they could automatically reject other combinations that relied on it. And since much of the data was symmetrical, the program could rule out mirror images of a configuration if it reached a dead end in one arrangement. Using these techniques, they reduced their search to about a billion configurations. They were joined in this effort by David Narvaez, a Ph.D. student at the Rochester Institute of Technology, who was a visiting researcher in the fall of 2019. Once they ran their code on a cluster of 40 computers, they finally had an answer: the conjecture is true in dimension seven. "The reason we succeeded is that John has decades of experience and insight into this problem and we were able to transform it into a computer-generated search," Heule said. The proof of the result is fully calculated by the computer, Heule said, in contrast to many publications that combine computer-checked portions of a proof with manual write-ups of other portions. That makes it difficult for readers to understand, he noted. The computer proof for the Keller solution contains all aspects of the solution, including a symmetry-breaking portion contributed by Narvaez, Heule emphasized, so that no aspect of the proof needs to rely on manual effort. "We can have real confidence in the correctness of this result," he said. A paper describing the resolution by Heule, Mackey, Brakensiek and Narvaez won a Best Paper award at the International Joint Conference on Automated Reasoning in June. Solving Keller's conjecture has practical applications, Mackey said. Those cliques that scientists look for to disprove the conjecture are useful in generating nonlinear codes that can make the transmission of data faster. The SAT solver thus can be used to find higher dimensional nonlinear codes than previously possible. Heule recently proposed using the SAT solver to tackle an even more famous math problem: the Collatz conjecture. In this problem, the idea is to pick any positive whole number and divide by 2 if it's an even number or multiply by 3 and add 1 if it's an odd number. Then apply the same rules to the resulting number and each successive result. The conjecture is that the eventual result will always be 1. Solving Collatz with the SAT solver "is a long shot," Heule acknowledged. But it is an aspirational goal, he added, explaining that the SAT solver might be used to resolve a number of less-intimidating math problems even if Collatz proves unattainable.

Five SCS Students Named 2021 Siebel Scholars

The Siebel Scholars Foundation has announced that SCS graduate students Brandon Bohrer, Rogerio Bonatti, Megan Hofmann, Hsiao-Yu Fish Tung and Lijun Yu are among the recipients of the 2021 Siebel Scholars award. Now in its 20th year, the program recognizes almost 100 students annually from the world's leading graduate schools of computer science, as well as business, energy science and bioengineering. It supports students in their final year of study. "This year's class is exceptional, and once again represents the best and brightest minds from around the globe who are advancing innovations in healthcare, artificial intelligence, the environment and more,” said Thomas M. Siebel, foundation chair. Bohrer, a Ph.D. student in the Computer Science Department, earned his undergraduate degree in computer science at CMU. He is studying formal verification of cyberphysical systems, developing mathematical methods that can guarantee the safety of software that controls physical systems. As a Ph.D. student in the Robotics Institute, Bonatti studies the intersection of machine learning theory and motion planning. Specifically, he creates methods for robust robot intelligence in real-world settings. Bonatti's work has been deployed in multiple applications, ranging from autonomous cinematography with aerial vehicles to drone racing. Hofmann is a Ph.D. student in the Human-Computer Interaction Institute. Her research focuses on the intersections of digital fabrication, healthcare and disability justice. She has developed new generative design frameworks that support makers in healthcare settings and has contributed to the burgeoning field of algorithmic machine knitting. A Ph.D. student in the Machine Learning Department, Tung is interested in building machines that can understand and interact with the world. Her research spans unsupervised learning, computer vision, graphics, robotics and language. Yu is a master's student in the Language Technologies Institute. His research focuses on improving public safety by understanding surveillance videos in 3D. His work has included a video-based traffic danger recognition system for detecting car crashes and alerting first responders, and a system for spotting suspicious behaviors in surveillance videos. For more on the Siebel Scholars, visit the foundation's website.

CMU's MoonRanger Will Search for Water at Moon's South Pole

Small, Speedy Rover Completes Preliminary Design Review

MoonRanger, a small robotic rover being developed by Carnegie Mellon University and its spinoff Astrobotic, has completed its preliminary design review in preparation for a 2022 mission to search for signs of water at the moon's south pole.Whether buried ice exists in useful amounts is one of the most pressing questions in lunar exploration, and MoonRanger will be the first to seek evidence of it on the ground. If found in sufficient concentration at accessible locations, ice might be the most valuable resource in the solar system, said William "Red" Whittaker, University Founders Research Professor in the Robotics Institute."Water is key to human presence on and use of the moon," explained Whittaker, who is leading development of MoonRanger. "Space agencies around the world are intent on investigating it."Whittaker and his team first approached NASA about using robots to search for lunar ice in 1996, and they will fulfill that vision a quarter century later by landing in 2022."This hasn't been quick or easy," Whittaker said. "It is stunning that after these many years we will have the first look."NASA will follow MoonRanger at a later date with its more capable Volatiles Investigating Polar Exploration Rover (VIPER), which will perform more rigorous and sustained exploration and scientific characterization of the ice.MoonRanger's lander will be the Masten Space Systems' XL-1, supported by the NASA Commercial Lunar Payload Services program. The rover will be one of eight science and technology payloads, which are supported by the NASA Lunar Surface Instrument and Technology Payloads program.The space agency said the payloads support its Artemis program, which aims to return U.S. astronauts to the moon in the coming years.Last month, reviewers determined the viability of the design for the rover and its mission. Lydia Schweitzer, a master's student in computational design who led the systems engineering team, said the two-day review involved more than 60 people — including veterans of the Apollo program and Mars rover project — who provided important suggestions and feedback.Schweitzer said the project involved a dozen faculty and staff members, as well as at least 90 students, including three semesters of enrollees in Whittaker's project course. Disciplines represented on the team comprise engineering, robotics, computer science, software engineering, human-computer interaction, architecture and design. The team also has taken advantage of a network of CMU alumni with expertise in space robotics to solve problems and optimize the rover's design.Even as MoonRanger takes shape, Whittaker and another student team continue to prepare for a 2021 mission in which a four-pound CMU rover called Iris and a CMU art package called MoonArk will travel to the moon on Astrobotic's Peregrine lander.MoonRanger features a number of technical innovations. About the size of a suitcase, it is designed to repeatedly explore at the rate of 1,000 meters per Earth day in both sunlit and dark conditions — unprecedented speed for a planetary rover. By contrast, a Chinese robot now on the far side of the moon has averaged less than a meter per Earth day.Unlike other rovers, MoonRanger doesn't carry isotope heating, so its battery and electronics will fail when night falls and cryogenic temperatures set in. Hence, the robot must accomplish its mission in less than the 14 sunlit Earth-days of the lunar month. It also is light and can't carry a big radio for communicating directly with Earth. It thus must return to the lander, with which it will establish short-range wireless communication so the lander's radio can relay the robot's findings to Earth."MoonRanger is going to be on its own for long periods of time," said David Wettergreen, research professor of robotics and co-investigator for the rover project, noting the rover will be out of touch with controllers on Earth as it does its explorationsThe mission was originally designed to demonstrate the capability of the rover. But NASA expanded it this spring to include the search for ice by adding its Neutron Spectrometer System (NSS) to MoonRanger. The NSS, developed by NASA Ames Research Center, measures the amount of hydrogen in the upper layer of the moon's soil, called regolith. Hydrogen abundance is correlated with the concentration of buried water ice. The NSS will be along for the ride, "ticking like a Geiger counter" when the rover passes over buried ice, then falling silent in bone-dry areas, Whittaker said.The rover's solar array is oriented vertically to capture the low sun angles experienced at the pole. The low sun also means that craters and dips cast deep, pitch-black shadows. The rover, therefore, will need to sense and navigate through darkness — another first. Since LIDAR sensors used commonly by Earth robots aren't yet available for small space rovers, MoonRanger achieves night vision by projecting laser line stripes ahead of it to model the darkened terrain, much as stereo cameras do in sunlight.Once it lands on the moon, MoonRanger will evaluate its driving, navigation and mapping capabilities in short jaunts near the lander. It will then attempt a series of distant treks to seek ice."If we could make a one-kilometer trek, we'd be very happy," Wettergreen said. "If we could do it twice, that would be amazing."Uncertainty is inescapable for a mission as ambitious as MoonRanger, Whittaker said."In the face of that, there is only the question of whether to do it anyway," he added. "This has all the elements of purpose, technology, exploration, science and fulfillment of vision. These leave no question about going for it and giving it our all."

AMD Provides Computing Resources To Support CBD's COVID-19 Research

AMD has donated access to high-end computing to two Computational Biology Department faculty members, Christopher Langmead and Min Xu, to assist them in research projects related to COVID-19. Langmead, an associate professor, is modeling the evolutionary landscape of coronavirus proteins, which will enable the design of a vaccine capable of protecting against both the SARS-CoV-2 virus that causes COVID-19 and other coronavirus species. Xu, an assistant professor, is using an artificial intelligence technique he developed to automate the large-scale analysis of SARS-CoV-2 images produced via Cryo-ET, a 3D visualization tool for studying subcellular structures and the virus' infection process in host cells. "AMD is proud to be working with leading global research institutions to bring the power of high-performance computing technology to the fight against the coronavirus pandemic," said Mark Papermaster, executive vice president and chief technology officer for AMD. Thus far, AMD has contributed 12 petaflops of total supercomputing capacity — either high-end systems or access to cloud-based clusters — to 21 institutions and research facilities.

CMU's Roborace Team Prepares for First Competition

An autonomous car programmed by a Carnegie Mellon University student team will race for the first time Sept. 24-25 when Roborace, an international competition for autonomous vehicles (AVs), begins its season on the island of Anglesey in Wales. In Roborace, each team prepares their own artificial intelligence algorithms to control their race car, but all of the teams use identically prepared AVs, compute platforms and venues. To prepare for this month's race, the CMU team spent the summer working on the fundamentals of driving and on building an optimal driving path. But this week was the first time they had the chance to run their computer code on a hardware simulator. "Our minimum goal is to be able to get the car to start driving crash-free for now," said Anirudh Koul, an alumnus of the Language Technologies Institute's Master of Computational Data Science (MCDS) program and the team's coach. But the CMU team, the first U.S. team in Roborace, is confident that it will soon be competitive with other teams that have previous experience in the racing series. "We are true to the CMU spirit — underpromise, overdeliver," Koul added. Roborace delayed its season because of COVID-19 concerns. Its new season now includes 12 races hosted over six events. No spectators will be allowed for safety reasons and the teams can operate their cars remotely if they can't travel to the race courses. The races will be livestreamed on Twitch. Times and dates are subject to change, but fans will be notified if they follow Roborace on Twitch. The CMU team was organized by members of the MCDS program and now includes eleven students from LTI, the Robotics Institute and the Information Networking Institute.

Machine Learning Models Identify Kids at Risk of Lead Poisoning

Proactive Approach Focuses Limited Health Resources Where Most Needed

Machine learning can help public health officials identify children most at risk of lead poisoning, enabling them to concentrate their limited resources on preventing poisonings rather than remediating homes only after a child suffers elevated blood lead levels, a new study shows. Rayid Ghani, Distinguished Career Professor in Carnegie Mellon University's Machine Learning Department and Heinz College of Information Systems and Public Policy, said the Chicago Department of Public Health (CDPH) has implemented an intervention program based on the new machine learning model and Chicago hospitals are in the midst of doing the same. Other cities also are considering replicating the program to address lead poisoning, which remains a significant environmental health issue in the United States. In a study published today in the journal JAMA Network Open, Ghani and colleagues at the University of Chicago and CDPH report that their machine learning model is about twice as accurate in identifying children at high risk than previous, simpler models, and equitably identifies children regardless of their race or ethnicity. Elevated blood lead levels can cause irreversible neurological damage in children, including developmental delays and irritability. Lead-based paint in older housing is the typical source of lead poisoning. Yet the standard public health practice has been to wait until children are identified with elevated lead levels and then fix their living conditions. "Remediation can help other children who will live there, but it doesn't help the child who has already been injured," said Ghani, who was a leader of the study while on the faculty of the University of Chicago. "Prevention is the only way to deal with this problem. The question becomes: Can we be proactive in allocating limited inspection and remediation resources?" Early attempts to devise predictive computer models based on factors such as housing, economic status, race and geography met with only limited success, Ghani said. By contrast, the machine learning model his team devised is more complicated and takes into account more factors, including 2.5 million surveillance blood tests, 70,000 public health lead investigations, 2 million building permits and violations, as well as age, size and condition of housing, and sociodemographic data from the U.S. Census. This more sophisticated approach correctly identified the children at highest risk of lead poisoning 15.5% of the time — about twice the rate of previous predictive models. That's a significant improvement, Ghani said. Of course, most health departments currently aren't identifying any of these children proactively, he added. The study also showed that the machine learning model identified these high-risk children equitably. That's a problem with the current system, where Black and Hispanic children are less likely to be tested for blood lead levels than are white children, Ghani said. In addition to Ghani, the research team included Eric Potash and Joe Walsh of the University of Chicago Harris School of Public Policy; Emile Jorgensen, Nik Prachand and Raed Manour of CDPH; and Corland Lohff of the Southern Nevada Health District. The Robert Wood Johnson Foundation supported this research.

Pandemic Spawns 'Infodemic' in Scientific Literature

New Policies, Technologies Could Help Make Sense of Flood of Information

The science community has responded to the COVID-19 pandemic with such a flurry of research studies that it is hard for anyone to digest them all. This conundrum underscores a long-standing need to make scientific publication more accessible, transparent and accountable, two artificial intelligence experts assert in a data science journal. The rush to publish results has resulted in missteps, say Ganesh Mani, an investor, technology entrepreneur and adjunct faculty member in Carnegie Mellon University's Institute for Software Research, and Tom Hope, a post-doctoral researcher at the Allen Institute for AI. In an opinion article in today's issue of the journal Patterns, they argue that new policies and technologies are needed to ensure that relevant, reliable information is properly recognized. The potential solutions include ways to combine human expertise with AI as one method to keep up with a knowledge base that is expanding geometrically. AI might be used to summarize and collect research on a topic, while humans curate the findings, for instance. "Given the ever-increasing research volume, it will be hard for humans alone to keep pace," they write. In the case of COVID-19 and other new diseases, "you have a tendency to rush things because the clinicians are asking for guidance in treating their patients," Mani said. Scientists certainly have responded. By mid-August, more than 8,000 preprints of scientific papers related to the novel coronavirus had been posted in online medical, biology and chemistry archives. Even more papers had been posted on such topics as quarantine-induced depression and the impact on climate change from decreased transportation emissions. Simultaneously, the average time to perform peer review and publish new articles has shrunk. In the case of virology, the average dropped from 117 to 60 days. This surge of information is what the World Health Organization calls an "infodemic" — an overabundance of information, ranging from accurate to demonstrably false. Not surprisingly, problems such as the hydroxychloroquine controversy have erupted as research has been rushed to publication and subsequently withdrawn. "We're going to have that same conversation with vaccines," Mani predicted. "We're going to have a lot of debates." Problems in scientific publication are nothing new, he said. As a grad student 30 years ago, he proposed an electronic archive for scientific literature that would better organize research and make it easier to find relevant information. Many ideas continue to circulate about how to improve scientific review and publication, but COVID-19 has exacerbated the situation. Some of the speed bumps and guard rails that Mani and Hope propose are new policies. For instance, scientists usually emphasize experiments and therapies that work; highlighting negative results, on the other hand, is important for clinicians and discourages other scientists from going down the same blind alleys. They also explore other ideas such as identifying the best reviewers; sharing review comments; and linking papers to related papers, retraction sites or legal rulings. The authors also focused on greater use of AI to digest and consolidate research. Previous attempts to use AI to do so have failed in part because of the often figurative and sometimes ambiguous language humans use, Mani noted. It may be necessary to write two versions of research papers — one written in a way that draws the attention of people and another written in a boring, uniform style that is more understandable to machines. Mani said he and Hope have no illusions that their paper will settle the debate about improving scientific literature, but hope that it will spur changes in time for the next global crisis. "Putting such infrastructure in place will help society with the next strategic surprise or grand challenge, which is likely to be equally, if not more, knowledge intensive," they concluded.

Google.org Donates $1 Million, Personnel to CMU's COVID-19 Forecasting Project

Full-Time Team Members Supplied Through Google.org Fellowship

Google.org has donated $1 million to Carnegie Mellon University to support COVIDcast, the university's effort to track and forecast localized COVID-19 activity nationwide. The company is also providing a full-time pro bono team of 12 Google.org fellows — including software engineers, a program and product manager, a user experience (UX) researcher, and a UX designer — to support CMU's Delphi Research Group for the next six months. Both contributions will help Delphi expand its efforts to provide both a geographically detailed view of various aspects of the pandemic and an early warning to health officials when the number of cases in a locale are expected to rise. Since April, Delphi has been producing a variety of COVID-19 indicators, such as self-reported symptoms, doctor visits for COVID-19 symptoms and the time people spend away from their homes. These indicators are publicly available in real-time on the group's COVIDcast visualization website and via online data access. Delphi has been producing forecasts since June and started sharing its forecasts with the U.S. Centers for Disease Control and Prevention in July. Delphi has worked with Google, Facebook and other companies to gather its disease indicators, which include medical tests, hospital admission and browser searches related to COVID-19. "Google's support has been — and will continue to be — absolutely vital to our COVID-19 response," said Ryan Tibshirani, who co-leads the Delphi group with Roni Rosenfeld. "We are indebted to Google for this support. This is the silver lining of this pandemic: You find out just how many people and organizations, such as Google, are truly generous, and want to help." "It's been remarkable to see leaders across different industries, sectors and areas of expertise come together during this critical time in our world's history," said Brigitte Hoyer Gosselink, head of product impact for Google.org. "Google.org is proud to support the team at CMU with this grant and our Google.org fellows as they work together to better understand and address the short- and long-term effects of the COVID-19 pandemic." Tibshirani, associate professor of statistics and machine learning, and Rosenfeld, head of the Machine Learning Department, created the Delphi group eight years ago to perform epidemic forecasting, most notably for seasonal influenza. Google has provided search data to support that effort for the last five years. This year, the CDC asked CMU to extend its tracking and forecasting to include COVID-19, and Google has continued to provide search data, as well as distributing a survey question asking users if they are experiencing symptoms associated with COVID-19. During the pandemic period, the Delphi Research Group has grown to more than 40 people, including faculty, students and staff as well as volunteer contributors from outside the university. Last year, the CDC named Delphi one of two National Centers of Excellence for Influenza Forecasting.

New Perception Metric Balances Reaction Time, Accuracy

Both Elements Are Critical for Applications Such As Self-Driving Cars

Researchers at Carnegie Mellon University have developed a new metric for evaluating how well self-driving cars respond to changing road conditions and traffic, making it possible for the first time to compare perception systems for both accuracy and reaction time. Mengtian Li, a Ph.D. student in CMU's Robotics Institute, said academic researchers tend to develop sophisticated algorithms that can accurately identify hazards, but may demand a lot of computation time. Industry engineers, by contrast, tend to prefer simple, less accurate algorithms that are fast and require less computation, so the vehicle can respond to hazards more quickly. This tradeoff is a problem not only for self-driving cars, but also for any system that requires real-time perception of a dynamic world, such as autonomous drones and augmented reality systems. Yet until now, there's been no systematic measure that balances accuracy and latency — the delay between when an event occurs and when the perception system recognizes that event. This lack of an appropriate metric as made it difficult to compare competing systems. The new metric, called streaming perception accuracy, was developed by Li, together with Deva Ramanan, associate professor in the Robotics Institute and principal scientist at Argo AI, and Yu-Xiong Wang, assistant professor at the University of Illinois at Urbana-Champaign. They presented it last month at the virtual European Conference on Computer Vision, where it received a best paper honorable mention award. Streaming perception accuracy is measured by comparing the output of the perception system at each moment with the ground truth state-of-the-world. "By the time you've finished processing inputs from sensors, the world has already changed," Li explained, noting that the car has traveled some distance while the processing occurs. “The ability to measure streaming perception offers a new perspective on existing perception systems,” Ramanan said. Systems that perform well according to classic measures of performance may perform quite poorly on streaming perception. Optimizing such systems using the newly introduced metric can make them far more reactive. One insight from the team's research is that the solution isn't necessarily for the perception system to run faster, but to occasionally take a well-timed pause. Skipping the processing of some frames prevents the system from falling farther and farther behind real-time events, Ramanan added. Another insight is to add forecasting methods to the perception processing. Just as a batter in baseball swings at where they think the ball is going to be — not where it is — a vehicle can anticipate some movements by other vehicles and pedestrians. The team's streaming perception measurements showed that the extra computation necessary for making these forecasts doesn't significantly harm accuracy or latency. The CMU Argo AI Center for Autonomous Vehicle Research, directed by Ramanan, supported this research, as did the Defense Advanced Research Projects Agency.  

Robotics Students Win Qualcomm Fellowship

The team of Xinshuo Weng and Ye Yuan, both Ph.D. students in the Robotics Institute, is one of 13 nationwide to win a 2020 Qualcomm Innovation Fellowship (QIF). The QIF program is unusual because it requires pairs of students to submit proposals. Qualcomm says this approach reflects its core values of innovation, execution and partnership. Finalists are selected by the company's top engineers, and students then must present their proposal to a panel of executive judges. The proposal by Weng and Yuan, "3D Multi-Agent Social Interaction Understanding and Diverse Future Behavior Forecasting," addresses how next-generation autonomous artificial intelligence systems — such as self-driving vehicles, delivery drones and assistive robots — can interact safely with agents such as people, animals and other robots. They will work to develop detection and tracking algorithms that can learn how these agents are related to one another using video and LiDAR point cloud data. Based on the system's knowledge of how the agents interact, the perception system will be able to predict what the agents will do a few seconds into the future. They plan to validate the system by testing it in crowded environments with a large number of people and robots. Qualcomm engineers will mentor Weng and Yuan as they develop their system during the yearlong fellowship.

Facebook Launches Competition To Improve Detection of COVID-19

Participants Will Use Symptom Data Gathered by CMU, UMaryland

Facebook is sponsoring a competition, the COVID-19 Symptom Data Challenge, in which participants will try to improve early detection and situational awareness of the outbreak using real-time symptom data that Carnegie Mellon University and the University of Maryland have been collecting with Facebook's help. CMU and Maryland are partners in the competition, along with the Joint Program in Survey Methodology and Resolve to Save Lives. The Duke Margolis Center for Health Policy is hosting the challenge. Phase I competitors will submit their novel analytic approach using public data from the Symptom Survey, which CMU's Delphi Research Group and the University of Maryland run. Facebook invites a sampling of its users each day to answer the survey questions about COVID-19-related symptoms and behaviors. The CMU and Maryland researchers run the survey and publicly share aggregate results. Facebook doesn't have access to individual responses. The Delphi group, which surveys U.S. users, shares its data publicly on its COVIDcast website. To date, it has collected more than 10 million U.S. survey responses. Applications for phase I are due September 29. Judges will select five semi-finalists based on validity, scientific rigor, impact and user experience, with each receiving $5,000. In Phase II, the finalists will develop prototypes that will be presented at an unveiling event. The first-place winner will receive $50,000 and the runner-up will receive $25,000. The winning design also will be featured on the Facebook Data for Good website. Learn more about the COVID-19 Symptom Data Challenge on the competition's website.