Where Industry Meets Innovation

  • Contact Us
  • sign in Sign In
  • Sign in with certificate
mit campus


Search News

  • View All
  • ILP News
  • MIT Research News
  • MIT Sloan Management Review
  • Technology Review
  • Startup Exchange

ILP Institute Insider

January 17, 2017

Optimizing the new networks

Asu Ozdaglar integrates social and economic factors into tools for understanding large-scale decentralized systems.

Eric Bender

In the age of mobile phones, new sources of information can help us reduce urban traffic congestion—if we have suitably deep insights into how to best exploit the information, says Asu Ozdaglar, MIT professor of electrical engineering and computer science.

Services such as Google Maps that track GPS-enabled phones “are providing information to commuters about a new route that has less congestion, with the goal of reducing travel time for their own users,” she says. “But the effect of these traffic information systems on overall congestion is not well understood. So one of my group’s efforts is to develop game theoretic models that describe how commuters make road choices for a given particular incentive or information structure, as a first step to optimizing those controls to reduce congestion and total cost in the system.”

Asuman Ozdaglar
Joseph F. and Nancy P. Keithley
Professor of Electrical Engineering
and Computer Science

That’s one example of Ozdaglar’s work in developing new models, mathematical tools and algorithms for optimization and control of today’s rapidly evolving large-scale decentralized systems.
Other examples come from her research on systemic risk in networks—looking at how local disruptions are amplified and affect a global supply chain.

As the 2008 financial crisis showed, “small individual shocks to some part of a system can propagate through the rest of the network, causing cascaded failures and even ultimately a complete meltdown of the system,” she says. “Our group develops models to understand which connectivity architectures minimize these effects and which amplify them. Having a systematic understanding allows us to design static and real-time intervention mechanisms to use resources and intervention strategies to contain the damage in the system.”

Ozdaglar emphasizes that many of these very large-scale systems can be seen as economic or social networks, in which individual human decision-making is crucial to the network operation.

The electrical power grid offers one case in point. “The grid is not just a technological network that delivers electricity to homes; you can easily view it as an economic network that constrains the flow of a good (electricity) from its competing producers to utility-maximizing consumers,” she says. “Models that actually incorporate individual decision making and its interaction with the underlying network structure are crucial in developing efficient optimization algorithms, and in designing incentives for ensuring more efficient operations.”

Given the breadth and depth of the factors that are weighing into these network connections, it’s not surprising that Ozdaglar cooperates closely with researchers in other disciplines. In her studies of systemic risk, for instance, she often partners with MIT economics professor Daron Acemoglu.

She also aims to expand her collaborations with industry. “I’m looking for partnerships in which we can have access to large-scale rich data sets, and use these algorithms to make predictions as well as decisions,” she says. These projects could range from access to data sets to designing new algorithms to placing student interns in corporations or inviting visiting researchers from industry into her lab.

Re-optimizing optimization

Until recent years, the traditional discipline of system optimization could assume that information was centralized, problem sizes were moderate to large, and there was a single well-defined optimization objective, she says.

Today, not so.

“Thanks to advances in sensor storage technologies and the development of online platforms, we’re now able to acquire much more large-scale data, whether it’s tick-by-tick data from the stock market or frame-by-frame high resolution images from surveillance systems,” she points out. “These data also are very large-dimensional, capturing very rich features.”

Crucially, these very large-scale networks are decentralized. “They connect many heterogeneous agents,” Ozdaglar says. “They connect many sensors that collect distributed information. And most importantly, they are not operated by a single administrative domain.”

Consider, again, the electrical power grid. Rather than being centrally managed, she notes, it is made up of competing power generators, system operators that manage transmission lines, local utility companies controlling distribution networks, and consumers who can decide when and how to run up their electric bills.

Additionally, Ozdaglar notes, underlying network structures help to deliver how these technological services operate, and how economic and social interactions take place in them. In the power grid, for instance, the underlying connectivity structures determine congestion levels and overall system performance.

“All these new considerations necessitate a new optimization paradigm that involves the development of optimization algorithms that are fast, scalable to process huge-dimensional data sets, can operate with local distributed information, and most importantly can recognize the role of social interactions with the underlying technological infrastructure,” she says. “Much of my work focuses on developing these fast and efficient algorithms.”

Deep into interdisciplinary research

Ozdaglar also directs the Laboratory for Information and Decision Systems (LIDS), one of four labs within the Electrical Engineering and Computing Sciences department.

“LIDS is a center of gravity within MIT, nationally and internationally in research on information and decision sciences and applications,” she says. “It’s a very cohesive and collegial research community where a lot of collaborations take place between principal investigators and other faculty. We still do a lot of research in the core disciplines, but we’ve also expanded into emerging interdisciplinary areas, such as the intersection of game theory, economic modeling and social and economic networks.”

In 2015 LIDS joined the newly launched Institute for Data, Systems and Society (IDSS), whose mission is to advance research and education at the intersection of decision systems, statistics and data science, and social science. IDSS also incorporates the Sociotechnical Systems Research Center, which focuses on large-scale data-driven projects to address sociotechnical problems, and a new Center for Statistics, which brings together research in the field and attracts additional faculty members.

“Interdisciplinary research is hard, because it takes a long time to even speak each other’s language, find problems that are mutually interesting and actually commit to working on these projects together, as opposed to acting as consultants,” she remarks. Given MIT’s vast breadth of expertise, she adds, doing these interdisciplinary efforts at MIT is compelling, and IDSS makes these collaborations more systematic by bringing researchers under one umbrella.

Ozdaglar stresses that IDSS also offers a major new framework for joint projects with corporations. While the new institute is grounded in fundamental research, its mission is to address complex societal problems.

“We have started efforts in five initial flagship areas: finance, urbanization, social networks, energy systems and health analytics,” she says. “We very much look forward to partnerships with industry.”