Kagan Tumer's Research Interests:

Associate Professor, Dynamics and Control
Mechanical Engineering Department (Dynamics & Control)
Oregon State University

Phone: (541) 737-9899
Fax: (541) 737-2600
E-mail: kagan.tumer@oregonstate.edu
URL: htp://engr.oregonstate.edu/~ktumer
Office: 426, Rogers Hall
Mailing address: Mechanical Engineering Department, 204 Rogers Hall, Corvallis, OR 97331-6001

Home     Research     Teaching     Students     Publications     Bio     CV

Motivation:

The number of large complex systems composed of many interacting subsystems has exploded over the last decade. Spurred by the ever increasing size, interconnectivity and complexity of systems on one hand and the miniaturization and affordability of computing power on the other, new paradigms to managing such systems are emerging. Coordinating thousands of subsystems in dynamic and stochastic environments, an idea that a mere decade ago would have been outlandish, is not only possible, but imperative today. Indeed, the technological bottlenecks today are due to the lack of mathematics and algorithms to manage and coordinate such systems rather than difficulties associated with building them.

My current research deals squarely with this issue. Specifically, I focus on ``collectives", a set of utility-based agents that optimize a system level objective through pursuing their own local objectives. The critical challenge in a collective is in determining the proper local objectives that when pursued successfully by the agents, lead to good system level behavior. To date, applications of my work have included:

  1. Coordinating multiple robots/Unmanned Aerial Vehicles (UAVs)
  2. Alleviating congestion in traffic problems.
  3. Routing power/data over a network;
  4. Controlling constellations of satellites; and
  5. Coordinating thousands of nano or micro computing devices.

Early and Current Work:

My early work focused on deriving a mathematical framework that quantified the benefits of classifier ensembles. I formalized the intuitive notion that combining classifiers is more beneficial when the classifiers are independent, and quantified the expected gains. The subtle but crucial impact of this work was in pointing out the importance of independent errors rather than independent classifiers, which led to a new and practical estimate for the Bayes error rate.

I have since moved on to the study of collectives, where component interactions are significantly more complex than a handful of classifiers with a common objective. My main tasks in this endeavor have been to identify and quantify the system characteristics required for achieving coordinated behavior through the interactions of multiple agents. Two such characteristics constitute the underpinnings of my work. The first is the concept of aligning the agent objectives with the overall system objectives and is akin to a company providing stock options to employees. The second is the concept of objective-sensitivity which measures how dependent an agent's objective is on the actions (states) of that agent. In many ways, this work can be viewed as ``inverse game theory''. Rather than studying the equilibrium points of a many player system with an already defined payoff structure, I focus on designing the payoff structures that will lead to (globally) desirable equilibrium states.

Long Term Directions:

My long term goals are to extend the scope of this work both in terms of the application domains and the underlying mathematical framework. I will highlight three application domains I intend to pursue and briefly describe the interesting aspects of viewing them as collectives and the unique theoretical challenges offered by each:
  1. Multiple robot/UAV coordination: This domain provides a large scale complex system control challenge in which multiple robots have to be controlled to achieve system-level goals. The appeal of this domain from a collectives perspective is that in addition to coordination, this domain requires system level recovery from faults (malfunctioning robots), system reconfiguration (redistributing tasks to functional robots), and operation under communication restrictions (robots in and out of contact with each other, human operators).
  2. Micro-nano control: This domain presents the most severe scaling challenge as thousands to tens of thousands of unreliable components need to be coordinated. This is a particularly interesting domain since most of the more traditional learning and control methods are difficult to apply as many of the assumptions on component specification and performance do not apply.
  3. Airspace Management: This domain combines parts of the previous two challenges. While micro-nano device domain focuses on controlling many simple devices and the multiple autonomous vehicle domain focuses on controlling a smaller number of sophisticated devices, the airspace problem presents the challenge of controlling a large number sophisticated components.
Pursuing all three domains is essential in ensuring that the principles of controlling large dynamical systems stays paramount in this work.

Home     Research     Teaching     Students     Publications     Bio     CV