LIDS Student Conference photos taken by Qingqing Huang, Andrew Mastin and Yuan Shen. Commencement
Reception photos provided courtesy of Andrew Mastin and Mitra Osqui. Pictures for A Celebration of Scholarship in honor of Dimitri Bertsekas taken by Yuan Shen. Portraits of Kimon Drakopoulos and Vahideh Manshadi taken by Jennifer Donovan. Portrait of Dimitri Bertsekas taken by Donna Coveney.
Massachusetts Institute of Technology
Laboratory for Information and Decision Systems
77 Massachusetts Avenue, Room 32-D608
Cambridge, Massachusetts 02139
http://lids.mit.edu/ send questions or comments to lidsmag@mit.edu
A Message from the Director
On behalf of the entire LIDS community, let me welcome you to the Fall 2013 volume of LIDS|ALL. Within these pages you’ll find interviews and news that provide a picture of the activities within our Lab and the people for whom LIDS serves as their MIT home. As you look through this issue, our website, our new Facebook page, and MIT news releases, I believe you’ll find that LIDS is not only a dynamic and fulfilling place for all who call it home, but also an intellectual community whose contributions to and leadership in the broad field of information and decision systems continues to grow, receiving significant recognition within MIT and throughout the world. We are delighted to share these glimpses into our professional home. In this issue you will find an interview with Rachel Cohen, one of the members of our outstanding administrative staff. You’ll also find interviews with two of our senior graduate students, Ammar Ammar and Kimon Drakopoulos, whose contributions to LIDS extend well beyond their own research to service to the entire LIDS community. Vahideh Manshadi, one of our cadre of superb postdoctoral researchers, is also interviewed in this volume, as are Professor Dimitri Bertsekas, LIDS Alumna Angelia Nedich, who is on the faculty of the Depart-
ABOUT LIDS
The Laboratory for Information and Decision Systems (LIDS) at MIT, established in 1940 as the Servomechanisms Laboratory, currently focuses on four main research areas: communication and networks, control and system theory, optimization, and statistical signal processing. These areas range from basic theoretical studies to a wide array of applications in the communication, computer, control, electronics, and aerospace industries. LIDS is truly an interdisciplinary lab, home to about 100 graduate students and post-doctoral associates from EECS, Aero-Astro, and the School of Management. The intellectual culture at LIDS encourages students, postdocs, and faculty to both develop the conceptual structures of the above system areas and apply these structures to important engineering problems.
ment of Industrial and Enterprise Systems Engineering at the University of Illinois at Urbana-Champaign, and Dr. Henrique Malvar, Distinguished Engineer and Chief Scientist of Microsoft Research and a member of the LIDS Advisory Committee. In addition, you’ll find information about honors and awards received by members of the LIDS Community You will also find information on and photos of many of the activities during this past year including the exceptional annual LIDS Student Conference.
As some who read this already know, I’ll be stepping down as Director of LIDS at the end of the 2013 calendar year. It has been and continues to be a pleasure and privilege to serve as LIDS Director, working with this community of exceptional individuals who continue to make LIDS an inspiring and energizing place to be. The future for LIDS is bright, and, as a member of the LIDS community, I look forward to participating in and witnessing the accomplishments yet to come.
Sincerely,
Alan Willsky, Director
The Language of Math
By Genevieve Wanucha
Dimitri Bertsekas’ forty-four years of contributions to areas such as optimization theory, data networks, dynamic programming, and largescale computation proves nearly impossible to measure. His 15 books take too long to summarize. But a few words capture his skill and passion: the language of mathematics. He ‘translates’ and ‘formulates’ and ‘expresses’ complex problems into a language of rules and numbers.
After spending his youth in Athens, Greece and earning his masters at George Washington University, Dimitri arrived at MIT in 1969 and completed his PhD thesis in systems science in two years. There, he witnessed a transition from a narrow focus on control theory to a much broader
focus on systems analysis and its set of applications, which included data networks and power, communication, and transportation systems. It was a time when computation was primitive. As Dimitri thinks back to them, he chuckles. “At that time, a computer had 64,000 bits of memory— that’s 64 kilobits, not megabits, not gigabits like we have now,” he says. “The idea of moving messages between computers with those capabilities was mind boggling.” Though he could have never predicted the complexity of today’s networks, he has always been able to see through to their underlying mathematical structure.
In the years that followed, computational resources expanded and data networks multiplied in com-
plexity. With this transformation, along came the observation that all networks similarly feature things—messages, cars, trains, electricity—moving in coordination along networks that cannot sustain a load beyond a certain point. Dimitri notes that when he attempts to optimize traffic or communication networks, he finds identical mathematics. “By studying the mathematical signatures of these problems, you tend to build an understanding uncluttered by their real world characteristics.”
Dimitri encountered his most sustained intellectual focus, optimization theory, during his first faculty position at Stanford University’s Engineering-Economic Systems Department. There, he met his greatest intellectual influence, mathematical scientist David Luenberger, who was one of the early pioneers of modern optimization. The then-emerging field involves making decisions as close to the best way possible in the face of uncertainty, such as calculating the most lucrative investment. “It was an exotic field when I started,” Dimitri says, “and now, optimization permeates just about every quantitative field. It has become the approach of choice in formulating problems.” Ever since, Dimitri has used the language of mathematics to create computational algorithms involving approximations, or attempts to get closer and closer to a solution reasonably fast.
most important discoveries and vivid memories. It was a night in the late 1970s, when Dimitri lived in Illinois. This was during his second faculty position at the Electrical Engineering Department at the University of Illinois at Urbana-Champaign. A relentless storm was flooding his basement, forcing Dimitri to empty the room with a bucket for hours on end. In between bucket-fulls, Dimitri sat at a small desk and attempted to solve a certain optimization problem. But he didn’t have a single book to consult. “So I decided I would start from first principles,” says Dimitri, “solving the problem as a layman would.” It was there in the wet basement, with a pencil and paper, that Dimitri came up with ‘auction algorithms,’ a completely new set of approaches to solving classic optimization problems of both theoretical and practical natures.
Optimization algorithm design, plus a coincidental rainstorm, brought Dimitri one of his
An algorithm is a well-defined, step-by-step procedure for calculations. Dimitri’s auction algorithms solve network flow, transportation, and assignment problems, which are fundamental optimization problems that involve agents and tasks that need to be distributed among the agents in a way that minimizes the cost of each task. For one simple example, a bunch of grocery trucks need to make deliveries to restaurants as soon as possible, so the time it takes to make a delivery rings up a cost. The best solution is whatever arrangement of trucks to destinations will achieve the lowest cost. Auction algorithms provide an intuitive way
to solve these kinds of problems because they are, as he once wrote, “couched in everyday experience.” As he sat in his basement thinking far outside the boundaries of theoretical literature, he produced an algorithm that operates like a real auction where people make competitive bids to naturally reach the best offers. These methods solved networktype optimization problems fast, and soon became a staple in the field. “Had I had books around me,” says Dimitri, “that idea would probably not have come to me. It was an important discovery, and happened under the most unpredictable of circumstances.”
In 1979, as Dimitri published his first auction algorithm, he began his career at MIT’s Laboratory for Information and Decision Systems (LIDS), where he still works as the McAfee Professor of Engineering. Dimitri soon hit it off with MIT’s Robert Gallager, who was then making a switch from information coding to data communications networks, essentially the progenitor of the Internet. This area offered Dimitri many opportunities to translate practical problems into the language of mathematics and implement optimization theory. It also provided fertile ground for the development of new computational models based on distributed large-scale computation.
At LIDS, where theory and practical application are closely linked, Dimitri is mostly methodologically oriented, but that doesn’t
mean his research stays in his books. “LIDS is not an Ivory Tower of ideas,” he says. For example, the government often requires a cost benefit analysis for projects such as building bridges or constructing highways. “The cost and benefits can only be assessed through predictions from mathematical models,” he says, “which will eventually be solved using optimization networks like the ones that we deal with
here on a more abstract level.” Students, he notes, are critical in this transfer of ideas from academia to real world problems, as they move from MIT to industry sectors.
Even as Dimitri uses the language of mathematics to express abstract ideas, he often engages with it on a literal level. “I have a kind of a pathology that few people have,” says Bertsekas. “I write a lot of books.” The 15 volumes, several of them influential textbooks, have given him ample opportunity to develop a very specific perspective on mathematical writing, which he notes is not the same as expository writing or really any other form of written language.
Writing, for Dimitri, consists of two written languages: English and math. English is rich and ambiguous; mathematical writing is contained and must always be unambiguous. And while expository and creative writing need excitement and varying sentence styles, “in mathematical writing, you want to make the interesting point stand out and everything
else boring, even repetitive” he says. Such a composition style respects the fact that math writing is never read in one go. To truly absorb a mathematical concept, one must read the same section over and over, and eventually connect different parts together. He has created a guideline, “Ten Simple Rules for Mathematical Writing,” aimed towards students who are writing papers and dissertations.
Books and teaching go hand-in-hand in Dimitri’s research life. In fact, they rely on the other. “I would not be able to write books without teaching,” he says, “because books develop from class notes, which tend to follow new research questions.” This cycle of research to teaching to books and back to research has itself come full circle. On May 17th, LIDS celebrated the publication of Dimitri’s most recent book, Abstract Dynamic Programming,
with a series of talks by his collaborators.
Dimitri often looks at the world through the lens of a camera. He says that mathematical research and artistic photography may seem completely different, but that both actually share the risk involved in trying to go beyond established areas into unexplored territory. “You don’t know what you are looking for while you’re looking for it,” he says. Over the years, he has built a large portfolio of travel photography inspired by the brilliant sunlight in Greece, the dramatic rock formations in California’s Death Valley, and the mysterious green fogginess of China’s rivers and mountains.
Dimitri observes that the Stata Center, the Frank Gehry-designed architectural marvel and home of LIDS, truly reflects MIT’s spirit of innovation in a way like no other campus building. It’s a “photographer’s paradise” because the sloping metal surfaces change the way it reflects light and colors across the entire day. Several of Dimitri’s Stata Center photographs hang along the hallways of LIDS. In one, the mirror-like metal siding of the Center distorts the image of a neighboring brick building, so the reflection appears as melting orange liquid. At first glance, you will think it is a painting. Dimitri has photographed it in the way he creates his algorithms, abstracting a complex structure in order to see it in a whole new light.
LOOKING BEYOND
By Rachel VanCott
In the early 1980s, Henrique “Rico” Malvar was just starting out in his academic career. He’d never lived outside of Brazil and had a wife, a young daughter, and an assistant professorship in electrical engineering. But he was about to head into the uncharted waters of a life in the United States.
To hear him speak, it’s hard to imagine that the energetic Chief Scientist of Microsoft Research and LIDS Advisory Committee member ever had doubts about his future or his ability. But before his success at Microsoft, before his influential work in signal processing, and before his graduate study at MIT, Rico had to face a critical question: “Can I really leave a developing country, with a family, and go to the U.S. – to the best school in the U.S. – and have a chance to be successful?” Ultimately, Rico took that chance, and the answer turned out to be a resounding yes.
Rico can trace his interest in electrical engineering all the way back to his first electronics construction kit, given to him by his father. Eight-year-old Rico was amazed and intrigued. As he grew, he worked on more complex projects, like amplifiers and sound systems. By the time he reached college he knew what he wanted to be. Electrical engineering was the natural choice for his undergraduate and masters degrees. After working for a few
years as an assistant professor, the time came for Rico to apply for a PhD. He sent out applications to a number of schools, but his acceptance letter from MIT was the first one to arrive, and that letter was the only one he needed to see.
Rico’s first impression of MIT was equal parts excitement and caution. He recalls two main thoughts: “I’m glad I’m here” and “this isn’t going to be easy.” Rico had always been the best at what he did. But MIT is full of bests: valedictorians, academic polyglots, and star performers from around the world. Rico had to learn to compete and thrive in MIT’s elite community. But even in tough times, Rico says, he saw his struggle as evidence that he’d come to the right place.
In Professor Dave Staelin’s group as part of MIT’s Research Laboratory of Electronics, Rico worked on signal processing problems. Specifically, he worked on signal processing as it applies to compression of digital signals. “We already had the vision,” Rico says, “It was clear that someday digital storage and communication would be cheap enough that it would be common for people to have pictures and music and videos all in digital format.” But digital media files can contain a large amount of information. For example, a single three-megapixel image has over three million picture ele-
ments, or pixels. Each one of those pixels also contains information about intensity and color components of that point in the picture. The number of pieces of information grows quickly. Instead of laboriously copying and transmitting each and every piece of information about a file, computer systems can compress image files for transfer. Rico’s work included the study of a type of compression known as transform coding, in which the image is analyzed and then reduced to a more simple, but accurate enough, version of itself.
At the time, that process was primarily accomplished by chopping an image up into smaller tiles, encoding each tile, transmitting the information, and then reassembling the image at the destination. But the process wasn’t perfect. When you break an image down into squares and encode each square separately, the edges might become slightly mismatched. Even a small mismatch becomes obvious in the reassembled image. Additionally, any noise in the signal or data loss may result in obvious artifacts--tiny blocks misplaced or missing in the final image. One of Rico’s lab mates, Philippe Cassereau, had proposed a method of resolving the problem by overlapping the blocks to eliminate mismatches and reduce the effect of noise or information loss. But Philippe’s work, though interesting, couldn’t be used for large matrices. Rico took the idea to the next level
and, as part of his thesis work, developed a practical and robust method of applying the idea. This method of overlapped orthogonal transforms soon came to be used in multimedia formats, Internet telephony, DSL modems, and other applications.
With his PhD from a world-class university in hand, Rico went back to Brazil and resumed teaching. He was promoted to associate professor and, if he had done nothing more than continue on his trajectory, he probably would have had a successful career there. But instead, he received a tempting phone call from Massachusetts. Some of Rico’s old lab mates had started a videoconferencing company called PictureTel; one of the first videoconferencing companies in existence. They called to tell him they were searching for a new director of research, and they invited him to apply.
Rico had always considered himself an engineer. He had always planned to be a professor or researcher. But now he found himself contemplating a transition out of academia and into industry, furthermore, into a management role, and another move for his growing family. He and his wife weighed the options, and decided to move back to the U.S. At PictureTel, Rico learned to direct his team and honed his understanding of how fundamental research turns into products and services. Along the
way, Rico started to pick up on little telltale signs of his success. Colleagues started seeking his advice and bringing him in on projects.
In 1997, Rico became interested in doing broader work and took an opportunity to join Microsoft’s Redmond, Washington research lab. He’s been with the company ever since, and risen through the ranks, from a Principal Researcher role, to Managing Director of Microsoft Research Redmond, and finally to his current role as Chief Scientist.
"The best motivation people can have is not money or promotion but really challenging their own mind."
the chance to pursue projects they find interesting, challenging, and promising.
“The best motivation people can have is not money or promotion but really challenging their own mind,” said Rico, “Prove to myself that I can do X. That’s when people do the most interesting things and have the biggest accomplishments.” At any rate, Rico says, Microsoft has so many products and developing technologies that almost any research focus will have some application or value to the company. Management does, of course, have some oversight. As projects develop, the management team prunes the output; directing more resources toward projects that hold the most promise and letting other projects take a back seat.
These days, Rico doesn’t have time for much research, though he publishes when he can –usually a few times a year. Primarily, he focuses on directing the research of others, encouraging his scientists and helping spur them on toward the development of the next big thing. Like any company, Microsoft has strategic goals. But instead of strictly directing work on specific products, Rico and the rest of the management team put the focus on advancing technology in general. They give their researchers
As a member of the LIDS Advisory Committee, Rico helps guide the direction of LIDS research with a much gentler nudge. He’s known Alan Willsky since he first took a class taught by the LIDS Director. When Alan invited Rico to be a member of the newly formed advisory group, Rico was happy to join. As part of the committee, he offers advice and opinions on the lab’s latest collaborations, research, and other kinds of development. The role also gives Rico a convenient excuse to visit MIT – a place that he loves but doesn’t see as often as he’d like.
Looking back on his career thus far, Rico is amazed to see the technologies that his work has touched. He sees the result of signal processing work everywhere: in the voice recognition and noise cancellation of cell phone software, in face and location recognition in image processing, in the motion detection and analysis of video game systems. The work that he’s done, in a small way, helped all those things become part of the modern world. Rico recommends that young researchers remember the bigger picture and understand why, at a high level, they’re doing what they do, even when working focused on intensely technical problems. “This is a personal choice, and not everyone needs to be this way, but try to do important stuff,” says Rico, “When you pick out problems, try to see beyond solving the math or beyond solving the algorithm. Understand how and why what you’re doing matters, and what might come of it.” It is a philosophy that has served Rico, and his groundbreaking work, very well.
SMART GRAPHS
By Genevieve Wanucha
Four years ago, when Kimon Drakopoulos was twenty-two, he left Greece to start his master’s degree at LIDS. It was an academic triumph, but just as much a personal one. In Athens, even though students are hardworking, Kimon sensed a widespread attitude that the important thing was to pass exams and get a good job. For Kimon, though, there’s something deeply unsatisfying about the idea of just getting through life – he wanted to do something more. LIDS, and its energetic and collaborative community, became the solution. “I think that’s a big part of why we are all here,” he says.
Intellectual ethos aside, Kimon works at LIDS for the chance to do something untraditional. He had studied computer vision in college, an active field focused on designing computer programs to replicate the brain’s ability to detect and categorize images and movements. The applications to such things as facial recognition and robotic control is interesting, “but the thing is,” says Kimon, “there have been so many people working on it for so many years that you need at least three years to just learn all the details. It takes a lot to do something significant in the field; or at least it would for people with an average mental capacity like me,” he jokes. “I wanted to work on something that hadn’t been explored so thoroughly yet, and make a bigger impact.”
faculty advisors, Profs. Asu Ozdaglar and John Tsitsiklis, makes this impact by modeling stochastic, or randomly evolving, processes on graphs. It is a different way of looking at how the world works. Graphs involve nodes connected with edges. Nodes can represent, for example, people, while the edges drawn between them reveal their connections. Using historical data, graphs can approximate things like the way information or disease flows through a social network.
One of the most significant future applications of Kimon’s work is contagious disease management. The question occupying him now has to do with finding the smartest way to make and distribute flu vaccines, which is limited by the high expense of manufacturing enough vaccines. With no idea about where the flu will spread, the best bet from a health policy perspective is to vaccinate everyone. But, Kimon points out, a smarter way to vaccinate would be to only give vaccines to people in the position to infect the most people.
“That’s my goal,” he says, “I want to find smart ways on graphs to produce the least amount of vaccine, but contain the epidemic fast.”
As a network scientist and second year PhD student, Kimon, along with his group and his
Kimon wants to use his graph models to decide whom to vaccinate in a certain group of people, given constrains on the budget for producing vaccines. In deciding which two people should get the vaccine, the important information is their neighbors. Kimon likes to
use the example of choosing between himself and a professor who will give a lecture tomorrow. “The professor would infect 300 people,” says Kimon, “and I would infect three.” Obviously, if the decision had to be made, the professor should get the vaccine. But there’s a twist. Consider the decision to vaccinate Kimon or a
woman who happens to be the professor’s wife. Suppose she has even fewer connections than Kimon, and would infect less people. However, one of the people she would infect would be her husband with the potential to infect the entire lecture hall. So, the vaccination decision depends not only on the immediate number of connections a person has, but also on the nature of their relationship to other people in the network. “That’s what we are trying to find out,” says Kimon, “who is central in the graph, in terms of their potential to infect people.”
Kimon is still tweaking the theoretical underpinnings of this graph. It’s not yet ready for real medical data. But when it is, one big challenge will come with understanding how to use these graphs without complete information about who has the flu or the nature of people’s connections. Kimon says that they will have to partly infer the structure of the epidemic and work out from there. Ultimately, this deeper understanding and ability to predict the scope of future infections would revolutionize the government’s readiness for public health emergencies.
Kimon worked as a data intern at LinkedIn during the summer of 2012, getting the chance to analyze loads of real social network data. From that experience, he says, “I am now pretty convinced that some of the current models are relatively realistic.” He hopes to begin working real data into such models by the end of his PhD work.
What’s even more exciting is that this graphing approach can be applied to any phenomena with “contagious” qualities, whether it is a spreading forest fire, disseminating blog post, viral tweet, or decisions within a social network about buying Androids vs. iPhones. An understanding of how these things evolve would be a goldmine to an endless array of policymakers, advertisers, government agencies and social scientists.
Contributing to research that can make the world a healthier and perhaps more predictable place is not the only thing that Kimon loves to do, though. In the spring and summer, he rides his Suzuki V-Strom to work during the week and into Boston on the weekends. Even though motorcycle riding in Cambridge is not the same thrilling challenge as it is in his hometown of Athens, it is safer and more enjoyable. He’s even ridden this on- and off-road vehicle across the entire United States during the summer of 2012.
In the winter, when traveling on his V-Strom is out of the question, Kimon lately discovered the gym as a true passion. He also cooks. In fact, he recently took a class in the techniques of French cuisine, hoping to finally break away from well-worn Greek recipes and try something new. “I’m a housewife during the winter and a motorcycle rider during the summer,” he says, with his usual mix of honesty and wit that charms his colleagues at LIDS.
experiences that have most influenced him at LIDS, one image pops into his mind: a small sign on his advisor’s desk that reads: Simplify. The motto belongs to John Tsitsiklis, a professor of electrical engineering and co-associate director of LIDS. “That defines his style,” says Kimon. “He’s great at finding the simplest and most intuitive way of tackling a problem, and until he does, things look very complicated.”
Kimon has come to value the power of simplicity throughout his field. Out of all the scholarly reading he’s done, he’s particularly impressed by the monumental work of P.D. Seymour, a mathematician now at Princeton University, who worked out major problems in the properties of graphs. For one specific property “there are 30-page-long papers trying to solve this thing,” says Kimon, “and he came up with a proof that was half a page. That’s something I admire.” Kimon keeps a certain phrase in mind: “take it back to the basics and everything will follow-at least hopefully.”
When Kimon thinks about the people and
Miracle Matches
By Genevieve Wanucha
‘Bipartite matching’ is the classic mathematical exercise of making pairs in messy bunches of things, such as matching job seekers to jobs, buyers to online ads, or online daters to each other. This method is at the heart of better market design. It’s also increasing the odds of medical miracles. At LIDS and MIT’s Operations Research Center (MIT-ORC), postdoc Vahideh Manshadi solves matching problems for a very specialized marketplace—one that ultimately enables more people to access life-saving kidney transplants.
More than 102,500 people are currently waiting for kidney transplants in the U.S. Only
about 17,000 people receive kidney transplants each year, and about 4,500 don’t get their transplant in time. In the face of kidney shortages, people often offer to donate an organ to a friend or family member in dire need. Unfortunately, about a third of the time, the donor and recipient are incompatible in either blood type or immune profile, resulting in a never-ending flow of hard-to-match patients. Yet, the methods designed by applied economic researchers like Vahideh provide the elegant solution of finding viable transplant matches between complete strangers.
“It’s an opportunity that might not be case for
other organs,” says Vahideh, referring to the revolutionary practice called ‘kidney exchange’ in which a living donor gives one of her two healthy kidneys to a stranger in order to secure another kidney for her loved one. More specifically, the simplest exchanges involve two or three patient/donor pairs, for all of whom the intended transplant is impossible, but for whom the patient in each pair could safely receive a kidney from a donor in another couple. These “pay-it-forward” surgeries enable two or three recipients to receive healthy kidneys.
The first exchange was performed in Korea in 1991, but it was almost a decade until they were successful in the United States, initially here in New England. Over the past eight or so years, hospitals such as Johns Hopkins and Mass General have performed increasing numbers of them.
Vahideh, a native of Iran, recent graduate of Stanford University, and happy new Bostonian, discovered her ability to optimize these lifesaving exchanges in the space of three hours at a 2011 workshop held by the Association for Computing Machinery (ACM) in San Jose. There, she met two of her future influences: Itai Ashlagi of the MIT Sloan School of Management and Harvard economist Alvin Roth (now at Stanford). The two researchers had planned the workshop to get the computer science community interested in improving the logic behind organizing kidney transplants.
Kidney exchanges that involve more than a simple two-pair swap require complicated matching algorithms, which have proven very effective in practice. Around the time of the San Jose meeting, the algorithm used by the National Kidney Registry was making news. In fact, it led to 175 successful transplants in 2011, including the 30-person chain of exchanges known as ‘Chain 124,’ the biggest exchange performed to date and featured in the New York Times: “60 Lives, 30 Kidneys, All Linked.”
Thanks to her work in the Stanford Operations Research group, Vahideh was used to thinking about matching problems with dynamic features, i.e. problems in which the settings change over time. Obviously, kidney exchanges are dynamic: incompatible pairs arrive and enroll in exchange centers gradually over time, and these centers need to make matches in a time-sensitive manner. The insight came easily to her. “I realized the core problem they were trying to solve was basically a dynamic matching problem,” she says.
When forming a closed loop of matches, all of the transplants have to be performed simultaneously, and this limits the number of pairs that can be in a cyclic exchange to 2 or 3. Now, real-life examples of kidney exchanges such as Chain 124 are proving the power of using a novel exchange method called a ‘chain.’ A chain kicks off with an altruistic donor (an
extra kidney not part of a pair), allowing the donations to continue on longer as the National Kidney Registry’s algorithm finds matches in the constantly growing pool of pairs. “However,” says Vahideh, “there have not been rigorous studies of how to design exchanges in dynamic settings.” With that in mind, Vahideh visited Itai Ashlagi at MIT and told him her
ideas about optimizing matches. It wasn’t long before she, Itai, and Patrick Jaillet, her current advisor at LIDS, formed a team uniting LIDS and MIT-ORC.
Arranging a kidney exchange involves a tricky trade-off. “You want to have as many transplants as possible, but you also want to keep people happy, meaning they don’t wait too long and get acceptable kidneys,” says Vahideh. Theoretically, the longer a center waits, the more data the algorithm can work with to produce matches. But waiting exacts a huge cost because people are on dialysis, and there’s more time for donors to renege on their offers.
For Vahideh, Itai Ashlagi, and Patrick Jaillet, their central question is the same question that all kidney exchange centers have: how long should we wait before arranging the exchange? One week? One month?
Recently, Vahideh and her colleagues took a hard look at the waiting problem to figure out when longer waiting periods to facilitate a kidney exchange would be worth the risk. The group analyzed an algorithm close to those used in practice today, running computer simulations with two years worth of confidential data from kidney exchange centers. This approach allowed them to consider antibody profiles and look at how the chances of finding compatible matches change as pairs join over time. They were able to confirm that certain policies are a good idea, and some are not.
For example, Vahideh found that if a center is limited to arranging 2-way exchanges and cannot afford to wait long enough, then an immediate (online) arrangement is a good solution. “If you need to match using only two-pair loops in an exchange, waiting one week won’t buy you much,” she says. “Even moderate waiting isn’t helpful for simple two-way matches.” Of course, if a center can afford to wait a couple months, then it would arrange significantly more transplants. But Vahideh’s group found that no exchange comes close to the benefit of dynamic ‘chains’ similar to ‘Chain 124.’ Even when the center cannot wait, these chains are extremely effective choices.
According to Vahideh’s work, the harder option is the best choice. While dynamic chains are difficult to initiate because the hospital must procure an extra kidney and operating rooms must coordinate surgery timings across the country, the transformative medical practice is worth it. “Complex structures have a huge benefit,” says Vahideh. “You have more freedom and finding matches is easier.”
The rare mix of theory and practice at LIDS has given Vahideh the chance to do something she never did at Stanford—working with real data in her optimization models. It’s possible for an optimization researcher to focus entirely on theoretical models that only simulate reality. But there’s something about incorporating real-word data that compels Vahideh. “What
I find fascinating,” she says, “is that you can do deeply mathematical probabilistic work, test it on real data, and see that it can be of use in real problems. Any recommendations we can make for kidney exchange programs matter.”
Indeed, improving the logistics of kidney exchange means perhaps endless solutions. Exchanges involve kidneys from living instead of deceased donors, which dramatically decreases the chances of organ rejection or early organ failure. It’s good for government finances, too. Medicare saves around $500,000 to $1 million each time a patient is able to stop dialysis after a live donor transplant.
For Vahideh, a certain spirit of freedom enables her work at LIDS. “My advisor, Patrick Jaillet, let me find something I really wanted to do,” she says. “It motivates me that you can go out and find something you care about.” The academic freedom, open and welcoming attitude among LIDS’ faculty, and endlessly available resources at surrounding institutions such as Harvard and Microsoft, are the keys to Vahideh’s productivity and success. It could also be that here in the home of the very first national kidney exchange, her theoretical work has never come this close to reality.
A New Kind of Puzzle
By Rachel VanCott
Angelia Nedich has been solving puzzles for as long as she can remember. As a child, she played puzzle games with her parents. As she grew, her skills developed and her interest in solving problems remained strong. By the time she got to school, others began to notice.
“I would have solutions before anybody could even think about it,” says Angelia, thinking back, “Eventually I guess the professors started recognizing that I had something. They kept encouraging me and when they make you feel good about yourself, you just keep going.”
Professors gave her books to study and urged her to compete and advance, and she kept doing what felt natural. She finished a bachelor’s degree in mathematics from the University of Montenegro, Podgorica, in 1987, a master’s in mathematics from the University of Belgrade, in Belgrade, Serbia, in 1990 and picked up a PhD in mathematics and mathematical physics from Moscow State University in Moscow, Russia in 1994. After that, she came to the US with her husband and had a son.
Soon, she started to look for a job that would help her take the next step toward the professional career in mathematics she’d always imagined for herself. But getting the type of position she wanted proved to be a new kind puzzle. “There was no way of getting to academia,” says Angelia. Though she technically had a PhD, her work in Moscow hadn’t required the breadth of
coursework that would have been expected in the US. So she decided to head back to school and get a second doctorate, in the hope that it would help her compete in the tough academic job market. She applied to several schools and counted herself lucky when she was accepted to MIT, where she worked at LIDS under the supervision of Prof. Dimitri Bertsekas.
She was back on the path to academia, but life was busier and more demanding than before. “Part of me was never really a truly typical kind of student at LIDS, because I was a mother,” says Angelia. On an average day, she’d push to finish as much work as she could during the traditional workday on campus, then head home to take care of her son and then, still later, put in a “third shift,” during which she toiled away on coursework and research.
Despite the long hours, Angelia says she never considered giving up or turning to a different career. Being a scientist or professor in an academic, intellectual community was what she wanted to do and where she wanted to be, without a doubt.
She got a taste of that community while working at LIDS, alongside people who tackled difficult problems in control, communications, optimization, and signal processing. In her courses, she encountered classmates who had not only an abstract, theoretical understanding of problems, but also the intriguing
ability to make guesses based on their knowledge of physical principles and deep familiarity with applications. They had a feel for the problems, while Angelia was predisposed to more of a pure, mathematical understanding.
“I had to stretch, because if I wanted to survive, it’s not the math that matters. It’s application domain,” says Angelia, “That’s the place to grow... You can learn the theory as much as you want but unless you find its use somewhere else, you will not be able to deeply appreciate its power and beauty.” So she did extra reading and cultivated her ability to see the problems in the way her colleagues saw them.
LIDS is a lab full of people who are devoted to the study of practical, applied problems. During her time at the Lab, Angelia conducted thesis work that formed the basis for the project that has, to her surprise, sustained her interest for the last seven years; a project that started out as a general model and has since been used to study more application domains than she ever expected: multi-agent networked system optimization.
Though some of the related groundwork for the model was laid during her graduate work, Angelia didn’t define the nature of the network model until the fall of 2006, when a friend and a colleague from MIT, LIDS Professor Asuman Ozdaglar, came to visit. Angelia had just left a four-year position in industry, and
was now working as an assistant professor at the University of Illinois at Urbana-Champaign (UIUC), where she is still working today.
During the visit, the two collaborated on the model development of a distributed resource allocation in a multi-agent system. A multi-agent system is a decentralized network of agents or processing nodes. Each individual has limited information--it knows only about its nearby neighbors. But the network as a whole is charged with solving a larger network-wide problem. This, Angelia says, was a more complex version of work she had done as a part of her thesis. As she explored the model, she found that by changing the restrictions and characteristics of the nodes or the precise goals of the network the model could be used to examine a variety of problems. “The directions and applications of that model have far exceeded any expectations I may have had when this started,” Angelia says.
lem more than once, Angelia says, students keep expressing interest in applying the multiagent system optimization framework to one new application domain after another. Recently, graduate students have applied it to distributed regression and estimation, sensor networks, smart-grids, and machine learning.
Angelia is happy at UIUC. At the Department of Industrial and Enterprise Systems Engineering,
“You can learn the theory as much as you want but unless you find its use somewhere else, you will not be able to deeply appreciate its power and beauty.”
she enjoys a collaborative, intellectual community that she describes as similar to LIDS. She’s grown accustomed to the open skies and gently rolling land of Illinois, too. She likes the relative calm after Boston’s frenetic pace, trafficjams and hard-to-find parking – the price of daily life in the big city.
She received an National Science Foundation Civil, Mechanical and Manufacturing Innovation Faculty Early Career Development grant to study the problem further, and she’s been working in that area ever since. Though she’s thought of moving on to the next prob-
The days are still long, though. Now, in addition to juggling research and family, she travels to many conferences, teaches and makes herself available to students whenever possible. Her work still fills the nooks and crannies of life, on nights and weekends. But she doesn’t mind much. It’s still the only job she could ever imagine doing.
Ranking the Possibilities
By Sarah Jensen
From earliest childhood, Ammar Ammar was fascinated by all things electronic and mechanical: cameras, radios, household gadgets, and appliances. “I took apart most of the small clocks in our house to try and see what made them tick,” he recalls.
Ammar’s mother and his father—an English professor in the Palestinian town where he grew up—championed his curiosity and encouraged him to explore the extensive library in their home. “My father answered the easy questions with anecdotes and the hard ones by pointing to his bookshelves,” says Ammar. When he was 6, Ammar was introduced to computers, and after a few years enjoying video games, he had the chance to write his first computer program, a trigonometric calculator, in the C programming language. Later, he designed a program that controlled lights through the printer port. “For the first time,” he says, “I realized that a passing thought could come to life through creating a program. It felt like magic.”
When he was in high school, Ammar’s mentor, a family friend from Wales who lived nearby, introduced him to the work of philosopher and mathematician Bertrand Russell. “She was inspired by him in college,” he says, “and she encouraged me to read his works. I started with his beautifully written autobiography and some of his popular essays, and found myself slowly drawn to his more abstract work. His
discussion of the problem of induction set me thinking seriously about such concepts.”
In 2005, Ammar followed his passions to MIT where he discovered a community equally inquisitive about how things work. A first-year probability course introduced him to Devavrat Shah, Jamieson Associate Professor in MIT’s Electrical Engineering and Computer Science department, who is based in LIDS. “I think I asked too many questions in that class,” says Ammar. “And Dr. Shah’s answers intrigued me.” Shah impressed upon Ammar the idea of research as a willingness to explore in the face of the unknown. “He taught me that there is always uncertainty at the outset, but we mustn’t put limits on our potential or the possibilities of what we might accomplish,” says Ammar.
Ammar received his undergraduate degree in computer science and electrical engineering in 2009, and still eager to face those unknowns, joined LIDS in his pursuit of a PhD in computer science. In his second year, he teamed with fellow student Srikanth Jagabathula, Sloan School associate professor Vivek Farias, and Shah, now his PhD advisor, to begin their innovative work on recommendation systems, the technology used on websites from Amazon to eBay to Yelp to help users identify the books and movies and restaurants they seek. “People are asked to make choices all the time,” says Ammar, “and we
felt the better we could make our recommendation systems, the faster people would be able to find exactly what they’re looking for and facilitate their decisions.”
Existing recommendation systems, Ammar and his team soon found, are seriously flawed. Relying as they do on a subjective scale—five stars for excellent, say, and one for poor—their results are a bit murky. One rater’s interpretation of three stars may not be the same as another’s. And individuals may rate the same item differently from one day to the next depending on their mood, leading to noisy data and oftentimes inaccurate recommendations.
A better way to capture customer sentiment, they realized, is through rankings, in which items are compared with each other rather than assigned individual scores. Ammar and his team set about creating a framework and a number of algorithms that accomplish just that, resulting in more precise and objective data. In 2011, they demonstrated that their approach is able—using a year’s worth of data on automobile sales—to predict preferences with 20 percent greater accuracy than existing algorithms.
The comparison method yields rich, accurate information, but the complex computation required to extract data from across all options and customer profiles is a challenge: As more choices are introduced, the number of possible
customer profiles grows exponentially. “If we ask people to compare coffee and tea, the result is two profiles: those who prefer coffee and those who prefer tea,” explains Ammar. “But if we add hot chocolate, there are six possible orderings.” Comparing four items yields 24 profiles, and if 10 items are compared, the rankings can be ordered in more than three million configurations.
That’s a lot of possibilities, and the team took one step back to realign their perspective. “We believed that fewer possible profiles actually exist in a population,” says Ammar. “And we realized it wasn’t necessary to explicitly keep track of all them.” Rather than tackle the dizzying task of analyzing millions of profiles, they discovered the data could be summarized—and yield the same robust results. Currently, they’re exploring ways to increase the resolution of that process through an algorithm that views the larger population as a collection of smaller populations. Narrowing the entire restaurant-going public to groups with shared taste profiles, for instance, would add a granularity to the rankings and enable potential consumers to more easily make choices based on the information provided by others in their subgroup.
Internet shoppers are another group that can benefit from the ranking system. The proliferation of online retailers has made Internet shopping more difficult than visiting a brick
and mortar store, says Ammar. “It takes time to search online in a brute force way to find the exact item you want.” Locating the perfect shirt, for example, could involve navigating among a dozen sites and sharing links with friends to gather their opinions—and could take hours. Further refinements to the group’s algorithm could streamline the virtual shopping experience, allowing customers to more quickly and accurately narrow their search and locate the ideal item based on price, style, and color—with only a few clicks of the mouse.
Ammar has begun work on additional practical applications of the technology, including gathering student preference information in order to place them in the class section most appropriate and convenient to them. Another potential application is a polling system to collect preferences of students and faculty, resulting in the ability to recommend activities, locations, and times that best match their leisure interests and schedules—and reserve a spot in their favorite section of the venue. Ammar’s work isn’t likely to stop there. He plans to spend the summer at Microsoft Research in Seattle investigating further possibilities in the area of public opinion polling.
He may step away from the lab from time to time and bring out glue and leather and indulge in his longtime avocation of vintage bookbinding, or join friends in a relaxing sail up the Charles River. But Ammar’s greatest
satisfaction, he says in a nod to physicist Richard Feynman, is the pleasure of finding things out. “I may approach a problem knowing very little,” he says. “It may not be clear in the beginning where a project will lead, but slowly, it begins to unfold.” He gives no little credit for his success to his LIDS colleagues and mentors. “In the LIDS community, we have the opportunity to share different perspectives, and that always helps clarify the challenges inherent in research,” he says. “Good ideas are always realized in collaboration.”
2013 LIDS Student Conference
ORGANIZING COMMITTEE
SPEAKERS
Student Conference
Chairs
Kimon Drakopoulos
Ermin Wei
Kuang Xu
Committee Members
Elie Adam
Ammar Ammar
Annie Chen
Austin Collins
HamzaFawzi
Christina Lee
Jenny Lee
Henghui Lu
Shen Shen
Omer Tanovic
Jianan Zhang
Giancarlo Baldan
Dr. Corinna Cortes
Wenhan Dai
Dr. Ron Dror
Hoda Eydgahi
Hamza Fawzi
Soheil Feizi
Prof. Sep Kamvar
Ali Kazerani
Christina Lee
Jeffrey Liu
Ali Makhdoumi
Andrew Mastin
Buddy Michini
Sidhant Misra
Marzieh Parandehgheibi
Jagdish Ramakrishnan
Luis I. Reyes-Castro
Hajir Roozbehani
Liangzhong Ruan
James Saunderson
Yuan Shen
Prof. Michael Stonebraker
Ramanarayan Vasudevan
Mengdi Wang
Andrew Young
BANQUET PERFORMERS
Qingqing Huang
Yola Katsargyri
Christina Lee
Prof. Alex Megretski
Shervin Mehryar
Noele Norris
Mitra Osqui
James Saunderson
Shreya Saxena
Omer Tanovic
Sze Zheng Yong
A Celebration of Scholarship
Honoring Prof. Dimitri Bertsekas on the occasion of the publication of his 15th book.
On May 17, 2013 LIDS hosted an event in honor of Prof. Bertsekas, who has made seminal contributions to many fields throughout his career thus far, including dynamic programming, data networks, optimization, and network algorithms.
The event featured talks by some of Dimitri’s collaborators: Profs. David Castanon (BU), Steven Shreve (CMU), Jonathan Eckstein (Rutgers), and Robert Gallager (MIT).
Sound Bites: Rachel Cohen
What do you do at LIDS?
I’m an Administrative Assistant at LIDS. I’m now assisting five LIDS professors: Sanjoy Mitter, Alan Willsky, Patrick Jaillet, Moe Win, and Yury Polyanskiy. I do some standard work for them, such as research reports and reimbursements. I also do some general administrative work for LIDS headquarters like keeping track of graduate student support and room assignments, and handling some web work for the LIDS Advisory Committee. I do a little of ev-
erything. I’ve definitely gotten pretty creative about keeping To-Do lists!
What do you like about working in LIDS?
It’s a wonderful place to work. I seem to have lucked into a place that is really good at understanding family needs. My son, Avi, is 10, and my daughter, Leah, is 13. My favorite aspect of LIDS is the quality of a work-life balance. In fact, I believe I’m one of the first people at MIT to have an official schedule allowing me
to work from home, so I value that opportunity. Overall, the people here at LIDS, and everyone I’ve worked with at MIT, have been wonderful, understanding, and appreciative.
What were you doing before you came to LIDS?
I’ve actually been at MIT for 17 years. I first worked in the Center for Theoretical Physics from 1994 to 2000, where I learned how to use LaTeX, a type-setting package written by a mathematician. When my daughter was born in 2000, I left for a year to be with her. After that, I came to work part-time for Sanjoy Mitter, who had just left his role as the co-director of LIDS. He needed someone who knew LaTeX, so that worked out fine. I started full-time at LIDS after about a year.
Have any of your prior experiences prepared you for your work at LIDS?
Absolutely. I majored in theatre arts design and technology at Cornell University. That’s all about organizing and pulling all the various elements of a theatrical production together and into performance and I really thrive on that. There is a definite parallel with the multiple tasks I handle at LIDS. Also, my dad is a physicist who spent 9 years at MIT. So, it’s natural for me to type mathematical papers and equations, as I am able to recognize the mathematical symbols, even if I don’t understand much of the specifics.
What
do you do when you’re not at LIDS?
Of course, I spend a lot of time with my kids and my husband Joshua, who is a professor of English literature at the Massachusetts College of Art. I have a lot of passions that come and go, but one is pretty consistent: I sew. I do upholstery, including pillows, curtains, and sofas for fun.
As my extended family has had more children, I’ve made quilts that I’ve illustrated for each new baby’s arrival with images of flowers, folk tales, Norse mythology, or Biblical stories. Sometimes, my husband draws the images and I sew them onto the quilts.
What are your plans for the future?
At the point when my kids don’t need me to drive them everywhere and can get themselves to school, I’ll consider getting more in involved in all the activities available at MIT. I plan on taking classes in pottery here. I’d also love to go to lectures on environmental science and energy. I read each issue of Energy Futures from MITEI cover-to-cover.
LIDS Awards & Honors
Congratulations to our members for the following achievements!
Awards
Amir Ali Ahmadi, Alex Olshevsky, Prof. Pablo Parrilo, and Prof. John Tsitsiklis received the 2012 INFORMS Computer Society Award, recognizing the significant contributions of three papers.
LIDS alum Amir Ali Ahmadi received the Simons Travel Award from the American Mathematical Society.
Former LIDS post-doc Anima Anandkumar was awarded a 2013 Microsoft Research Faculty Fellowship.
Prof. Asuman Ozdaglar was the inaugural recipient of the Steven and Renee Finn Innovation Fellowship, given in recognition of Asu’s accomplishments and her research push into new areas of great potential.
Prof. Pablo Parrilo was a recipient of the EECS Faculty Research and Innovation Fellowship. This is given to senior faculty members in recognition of outstanding research contributions and international leadership.
Prof. Yury Polyanskiy was selected for a CAREER Award by the National Science Foundation for his proposal titled “Information Theory Beyond Capacity”.
Honors
LIDS alum and former EECS and LIDS faculty member Nils R. Sandell, Jr. was named Director of DARPA’s Strategic Technology Office.
Prof. Devavrat Shah was named to the Advisory Board of Compass Labs in March 2013.
Prof. John Tsitsiklis was elected Chair of the Council of the Haropio University, in Athens, Greece.
Ermin Wei was selected as one of MIT’s Graduate Women of Excellence.
LIDS Director Prof. Alan Willsky delivered this year’s William Gould Dow Distinguished Lecture at the University of Michigan. This lectureship is the highest honor offered by the Department of Electrical Engineering and Computer Science at the University of Michigan.
LIDS Director Prof. Alan Willsky also gave the Dean Lytle Electrical Engineering Endowed Lecturer at the Department of Electrical Engineering and Computer Science at the University of Washington. The lecture series is the department’s premiere annual event, fea-
turing internationally renowned researchers in the field of communications and signal processing. To see videos of these lectures, visit the LIDS Facebook page (www.facebook.com/lidsmit).
Kuang Xu, a student supervised by Prof. John Tsitsiklis, was awarded a 2013 Claude E. Shannon Research Assistantship.
Paper Awards
Amir Ali Ahmadi, together with MIT coauthors Anirudha Majumdar and Russ Tedrake, received the Best Paper Award at the 2013 IEEE International Conference on Robotics and Automation.
Prof. Eytan Modiano, LIDS alum Krishna Jagganathan , and their collaborator at Qualcomm received the IEEE WiOpt 2013 Best Paper Award.
LIDS graduate Kazutaka Takahashi and his co-authors received the Best Paper Award at the 2012 joint International Conference on Soft Computing and Intelligent Systems at the International Symposium on Advanced Intelligent Systems held in Kobe, Japan.
Prof. John Tsitsiklis and his student Kuang Xu received the Best Paper Award at the 2013 ACM SIGMETRICS, where Kuang also received the Kenneth C. Sevcik Outstanding Student Paper Award for the same paper. This is the first time that both awards were given to
the same paper at SIGMETRICS since its inception in 1973.
Yuan Zhong, co-advised by Profs. Devavrat Shah and John Tsitsiklis, received the Best Student Paper Award at the Sigmetrics 2012 conference.
LIDS Seminars 2012-2013
Weekly seminars are a highlight of the LIDS experience. Each talk, which features a visiting or internal invited speaker, provides the LIDS community an unparalleled opportunity to meet with and learn from scholars at the forefront of their fields.
The Stochastic Systems Group seminar schedule can be found at: http://ssg.mit.edu/cal/cal.shtml
Listed in order of appearance.
Nicola Elia
Iowa State University
Electrical and Computer Engineering
Alon Orlitsky
UC San Diego
Electrical and Computer Engineering; Computer Science and Engineering
G. David Forney, Jr.
MIT
Electrical Engineering and Computer Science; LIDS
Mihailo Jovanovic
University of Minnesota
Department of Electrical and Computer Engineering; Control Science and Dynamical Systems Center; Digital Technology Center
Murat Arcak
UC Berkeley
Electrical Engineering and Computer Science
Nuno Martins
University of Maryland
Dept. of Electrical and Computer Engineering; Institute for Systems
Research
Alex Dimakis
University of Southern California
Dept. of Electrical Engineering - Systems
Piotr Indyk
MIT
Computer Science and Artificial Intelligence Lab; Theory of Computation Group
Vincent Poor
Princeton
Electrical Engineering; Dean of the School of Engineering and Applied Science
Pierre Moulin
University of Illinois at Urbana-Champaign
Dept. of Electrical and Computer Engineering
Jessy Grizzle
University of Michigan
Electrical Engineering and Computer Science; Control Systems Laboratory
Alexander Barg
University of Maryland
Dept. of Electrical and Computer Engineering; Institute for Systems Research
Gregory Wornell
MIT
Electrical Engineering and Computer Science; Research Laboratory for Electronics
Maxim Raginsky
University of Illinois at Urbana-Champaign
Dept. of Electrical and Computer Engineering
Sanjay Lall
Stanford
Aeronautics and Astronautics; Information Systems Laboratory
Sergio Verdú
Princeton
Electrical Engineering; Information Sciences and Systems; Program in Applied and Computational Mathematics
Maryam Fazel
University of Washington
Electrical Engineering
Bobak Nazer
Boston University
Electrical and Computer Engineering; Information Sciences and Systems
