Research Projects

Community Sense and Respond
Andreas Krause, Mani Chandy, Rob Clayton, Tom Heaton et al.
Can one use cell phones to detect earthquakes?

We tackle a fundamental question in cyber-physical systems: What is the ideal structure of systems that detect critical events – such as earthquakes – by using data from large numbers of sensors held and managed by ordinary people in the community? The approach is to develop theory about widely-distributed sense and respond systems, using dynamic – and possibly unreliable – networks using sensors and responders installed and managed by ordinary citizens, and to apply the theory to problems important to society, such as responding to earthquakes.
  • Project website of the Community Seismic Network
  • An IPSN 2010 paper with Matt Faulkner and Daniel Golovin on distributed online sensor selection [pdf]
  • An IPSN 2008 paper with Eric Horvitz, Aman Kansal and Feng Zhao on Community Sensing for urban traffic prediction from GPS devices [pdf]
Adaptive Submodularity
We are interested in practical algorithms with theoretical guarantees for challenging active learning and experimental design problems. We recently introduced a structural property, adaptive submodularity, which generalizes the classical notion of submodular functions to adaptive policies. We show that if a problem satisfies this property, an efficient, adaptive greedy algorithm guarantees near-optimal policies. We further show that this property arises in diverse applications including active learning, Bayesian sequential experimental design and adaptive viral marketing.
  • See our NIPS 2010 paper with Daniel Golovin and Deb Ray using adaptive submodularity for noisy active learning and nonmyopic value of information [pdf]
  • See our COLT 2010 paper with Daniel Golovin introducing adaptive submodularity [pdf]
Online
Which ads should we display in sponsored search in order to maximize our revenue? How should we dynamically rank information sources to maximize value of information? Which sensors in a sensor network should we activate when in order to best monitor some unknown phenomenon over time? These and other problems can be formalized as bandit problems optimization with complex, structured decision spaces. We are developing novel algorithms and analyze them in the no-regret model. We evaluate our approaches on several real-world information gathering problems.
  • Our ICML 2010 paper with Niranjan Srinivas, Sham Kakade and Matthias Seeger on Gaussian process bandit optimization [pdf]
  • Our NIPS 2009 paper with Matt Streeter and Daniel Golovin on learning to rank for diversity [pdf]
submodularity
Convex optimization has become a main workhorse for many machine learning algorithms during the past ten years. When minimizing a convex loss function for, e.g., training a Support Vector Machine, we can rest assured to efficiently find an optimal solution, even for large problems. In recent years, another fundamental problem structure, which has similar beneficial properties, has emerged as very useful in a variety of machine learning applications: Submodularity is an intuitive diminishing returns property, stating that adding an element to a smaller set helps more than adding it to a larger set. Similarly to convexity, submodularity allows one to efficiently find provably (near-) optimal solutions. We are studying submodularity and its implications for challenging sparse representation, inference and learning problems.
  • Tutorials (with Carlos Guestrin): ICML 2008, IJCAI 2009
  • Matlab/Octave Toolbox for submodular optimization and a JMLR paper about it [pdf]
  • See our NIPS 2010 paper with Peter Stobbe on efficiently minimizing decomposable submodular functions [pdf]
  • An ICML 2010 paper with Volkan Cevher on a connection between submodularity and sparse representations [pdf]
  • An ICML 2010 paper with Ryan Gomes on large scale budgeted nonparametric learning [pdf]
  • Our KDD 2010 paper with Jure Leskovec and Manuel Gomez-Rodriguez on a structure learning problem of inferring influential news websites and blogs [pdf]
Interaction Modeling
We are interested in rich probabilistic models for modeling interaction in navigation and activity recognition. As an example, we study the safe navigation of a mobile robot through crowds of dynamic agents with uncertain trajectories. Existing algorithms suffer from the "freezing robot" problem: once the environment surpasses a certain level of complexity, the planner decides that all forward paths are unsafe, and the robot freezes in place (or performs unnecessary maneuvers) to avoid collisions. Our key insight is that dynamic agents solve the frozen robot problem by engaging in "joint collision avoidance": They cooperatively make room to create feasible trajectories. We develop IGP, a nonparametric statistical model based on Dependent Output Gaussian Processes that can estimate crowd interaction from data. Our model naturally captures the non-Markov nature of agent trajectories, as well as their goal-driven navigation.
Optimizing Sensing
We identify a key structural property of many observation selection problems: submodularity, an intuitive diminishing returns property. Exploiting this property allows to obtain efficient approximation algorithms with provable guarantees, as well as drastically speeding up algorithms. We use our algorithms for applications such as sensor placement in drinking water distribution and environmental monitoring using robotic sensors.

  • A short overview paper with Carlos Guestrin on this topic that appeared as Cover Feature in IEEE Computer Magazine.
  • Our JMLR paper with Carlos Guestrin, Brendan McMahan and Anupam Gupta on the Saturate algorithm for robust submodular optimization [pdf]
  • Our JAIR paper with Carlos Guestrin on VOIDP for optimal value of information in chain models [pdf]
  • Our IJCAI 2009 paper with Amarjeet Singh and William Kaiser on adaptive informative path planning [pdf]