Current Research

My research is centered around networked systems and control. I work on developing efficient distributed multi-agent coordination algorithms mainly in three areas of

  • Consensus,
  • Optimization,
  • Cooperative robotics.


Input Average Consensus is a canonical problem that often appears in the coordination of multi-agent systems. Given a set of input signals, one per agent, this problem consists of designing a distributed algorithm that allows agents to asymptotically track the average of the inputs using only information from the neighbors. My research on average consensus algorithm design is focused on offering a solution that satisfies also practical requirements such as consistent response over different possible communication topologies, control over rate of convergence and time of arrival for individual agents, handling of limited control authority and preservation of privacy of local inputs as well as agreement states of agents. The first three issues are of importance in physical processes such as coordination of multi-vehicle systems. Privacy preservation is crucial in applications involving sensitive data.

Practical distributed in-network optimization algorithm design

The dramatic increase in available modern datasets and accompanying decentralized storage across different entities is driving the development and use of parallel and distributed algorithms for solving large-scale optimization and machine learning problems. Demand for distributed optimization algorithms is also driven by the surge of networked cyber physical systems, which rely on smart and optimal coordination of numerous subsystems to solve important socio-economical challenges, such as smart grid operations, smart transportation, and smart water irrigation to name just a few network operations that are envisioned to help us efficiently manage our resources. Although many of the problems can be cast as in-network convex optimization problems, these distributed operations are expected to take place over a medium characterized by the interconnection of multiple components via channels constrained by noise, security requirements, synchronization issues, and limited resources for execution. The challenge becomes not only to design algorithms that enable agents to compute their respective optimal operating point in a distributed manner, but also to do this in the presence of evolving network-induced constraints.
We work on devising holistic solutions for distributed convex optimization problems over networked system that takes into account operational constraints. Our approach blends control theoretic concepts with tools from parallel and distributed numerical computing paradigms. A control theoretic approach has the advantage of facilitating the characterization of properties such as speed of convergence, evaluating transient behavior and tracking performance for time-varying problems, managing disturbance rejection and robustness to uncertainty, as well as conducting observability tests for privacy preservation evaluations. For example, using an event-triggered control concept, we have been able to devise distributed unconstrained convex optimization algorithms in which each agent autonomously decides to trigger information to its neighbors only when it is necessary for the algorithm’s convergence and stability, resulting in an active asynchronous, energy-efficient, and robust communication strategy. We continue to work on expanding our results to constrained optimization problems. Concurrently, we are using concepts from observability theory in control to develop systematic tools that evaluate privacy preservation of distributed optimization algorithms and are devising transformation methods to induce intrinsic privacy preservation in these algorithms. Transparent privacy attributes will increase the adoption of distributed optimal network operations by private contributors. For robustness characterization against communication link noise we plan to use input-to-noise Lyapunov analysis. On a different subject, we also work on laying down foundations to develop central and distributed optimization algorithms for problems with time-varying cost and constraints. In such problems the objective is to track the time-varying optimizer. Such algorithms will facilitate dynamic optimal decision-makings in network operations.

Cooperative mobile robots operations

qrcode Distributed localization for multi-vehicle networks: significant growth in autonomous networked multi-vehicle applications in recent years drives research on design of optimal distributed strategies for mobile multi-agent systems. The success of tasks undertaken by these systems hinges on accurate localization of each of the vehicles involved. We have successfully devised centralized-equivalent distributed cooperative probabilistic localization using vehicle-to-vehicle measurements for GPS denied environments. Our current research direction is driven by the following observation: in practice, normally the motion control tasks for mobile multi-vehicle systems are designed with the assumption that there is a distributed algorithm in place providing accurate localization information. Although extensive simulations and experimental demonstrations are used to evaluate the behavior of the combined distributed localization and distributed control algorithms, such validations are heuristic with no guarantees for optimality or robustness. We plan to work on designing integrated distributed strategies for simultaneous control and localization for multi-vehicle systems. More specifically, we intend to design distributed algorithms that jointly optimize both the online control objectives and reduce the error in localization. Such formulations result in in-network static or dynamic optimization problems. We plan to leverage our group’s expertise in design of distributed optimization algorithms to solve these problems.

Slides: Take a look at these slides or this poster.

qrcode Robotic Lab: our robotic test bed consists of 5 Turtlebot robots, overhead cameras and RFID tags and antenna. We use our robotic test bed to test and evaluate our distributed algorithms for mobile agents. Our lab is also engaged in RFID tag localization and target detection experiments. Our testbeds run on Robotic Operating System (ROS) software, which is a framework for robot software development, providing operating system-like functionality on a heterogeneous computer cluster. If you are interested in working in my robotic lab (high school students, undergraduates and graduate students), please contact me at

qrcode qrcode qrcode

Other Research Activities

Controller Design for Systems with Bounded Actuators

Designing high performance controllers for systems with bounded actuators is theoretically challenging and, since actuator saturation is ubiquitous, it is critical for practical applications. For a long time ad-hoc but intuitive techniques were used. Over the last decade and a half, a variety of techniques with rigorously established stability and performance properties have been developed. A common method to deal with input saturation is the so-called Anti-Windup (AW) compensation method. Typically, AW is a two-step procedure, in which the original linear controller is designed without considering the input saturation. Then, the system is augmented with an AW protection loop to handle occasional saturation. In the traditional AW policy, the AW compensator is activated upon the occurrence of saturation. In my research, I developed a supporting theory to the long-observed phenomenon that sometimes linear controllers are best `left alone' to deal with small amounts of saturation and the immediate assistance from compensators can sometimes degrade small signal performance. My design activates AW gains when the saturated controller causes undesirable performance degradation. I also developed efficient AW scheduling schemes which allow systems to use more aggressive gains in lower levels of saturation. qrcode

Another way to deal with saturation is to explicitly take this nonlinearity into account in the controller design stage. However, this approach results in conservative controllers which work poorly in the small signal region. To reduce such conservatism, I developed a continuous family of controllers with increasing levels of aggressiveness or performance which will avoid saturation for a given bound on the worst case exogenous input.

Dilated Matrix Inequalities to Reduce Conservatism in Multi-objective and Robust Synthesis Problems

In this work, I presented a new variation of dilated matrix inequalities (MIs) for Bounded Real MI, invariant set MI and constraint MI, for both state and output feedback synthesis problems. In these dilated MIs, system matrices are separated from Lyapunov matrices to allow the use of different Lyapunov matrices in multi-objective and robust problems. To demonstrate the benefit of these new dilated MIs over conventional ones, they were used in solving controller synthesis problem for systems with bounded actuator in disturbance attenuation, and in robust H2 and H-infinity control synthesis. It show that for multi-objective and robust control design problems, the new form of dilated MIs are guaranteed to achieve performance measures which are less than or equal to the values achieved by conventional method. Also, in this research, I drew parallel between the proposed dilated matrices and the results that can be obtained using Full Block S-procedure, normally used in LPV problems.

Effects of Human Pilot Energy Expenditure on Pilot Evaluation of Aircraft Handling Qualities,

I proposed a methodology to model and predict the human pilot rating of a typical aircraft for flights through turbulence. The work was focused on developing a specific type of correlation between the pilot rating (PR) and the proposed pilot energy related parameters. The research involved a suitable pilot/vehicle analysis in addition to a computer simulation of an aircraft flying through a general turbulence. The `Beech' model 99 was used as the test aircraft due to availability of the required data and its known behavior in different flight conditions. The pilot model used in the simulation was based on the `Crossover' pilot model. The so-called `Neal-Smith' (NS) criterion was used to indicate pilots opinion.

Motion Wash-out Filters in Flight Simulators

I studied and investigated the Motion Washout Filters in literature which are used to generate near actual motion feelings in human pilots flying flight simulators. The Motion Washout Filters use concepts of human motion perception to generate the expected acceleration feeling in the human pilots using a combination of projection of the gravity and the accelerations generated through the limited motion envelope of the tion Washout Filters for Flight Simulators, typical motion systems of flight simulators. Thesefilters have efficients that can be adjusted to alter the motion response of the simulator. The work focused on selecting the coefficients considering the capabilities of the proposed motion system. The results were simulated in Matlab/Simulink.