Research and Interests
I am interested in networked systems – particularly how to design techniques that enable efficient coordination and operation of large-scale systems.
Our society and lives are increasingly interconnected – from communication networks, to robot swarms, to social networks and even networks formed by the emerging sharing economy (e.g. Uber, Airbnb), where humans are decision makers. Information and communication technology has led to the evolution of systems where devices embedded with sensing, computation and decision making abilities interact with the environment and cooperate among themselves to make decisions and coordinate system operation. Complex and spatially distributed systems like a network of drones and the power grid are not only expected to operate in uncertain, resource-constrained environments; but also need to be resilient and robust to vulnerabilities and attacks that can have widespread impact.
My research program – at the intersection of network science, limited-communication decision-making, distributed algorithms and security – centers on the design and implementation of techniques to efficiently coordinate the operation of spatially distributed large-scale networked dynamic systems. I work on communication and coordination challenges in Cyber-Physical Systems and other networked autonomous systems such as multi-robot systems and networks of smart devices. My research program can be summarized into the following thrusts:
Communication-Efficient Distributed Optimization in Large-Scale Systems
As robot swarms increasingly get deployed for tasks such as surveillance in disaster struck areas or even remote (arctic) areas with little to no communication infrastructure, the design of communication-efficient coordination techniques is central. The broad objectives are to design and characterize the performance of communication-efficient optimization techniques in fully distributed systems, where a number of challenges in communication arise. The general approach is to optimize information flow for distributed coordination and decision-making. The techniques being developed are especially vital in coordinating aerial robots outdoors, where the supporting communication infrastructure is limited; and underwater autonomous systems, where communication channels are known to be extremely weak. Techniques being developed here are equally important in the context of efficiently using wireless resources as we record explosive growth in the volume of connected devices.
Cognitive and Opportunistic Coordination for Resource Management in Networks
In networks of smart devices, for example, local factors such as devices’ available power, information processing and data storage abilities, and global factors such as the channel bit rates and available communication bandwidth determine its optimal operating point. The goal, in this research direction, is to embed cognition and intelligence to eliminate redundant packets of information in the network, and efficiently use network resources in coordinating its operation. The approach, at a high level, is to exploit the interplay between local cooperation, global constraints, data aggregation techniques and communication protocols using realistic communication models in cyber-physical networks and systems of connected devices. As captured by the Figure above, the idea is to design a coordination framework that learns, and makes sensing and transmission decisions based not only on each device’s local observations, but also on the history of decisions made by connected or neighboring devices as well local energy constraints and channel conditions. The interaction of efficient protocol design and state of communication channels alongside information processing in the cloud, for example, inform how data is aggregated and transmitted thereby efficiently using the network resources.
CPS Security and Network Protection
As CPSs and other networked systems become more integral in commercial and other critical infrastructure, they increasingly become vulnerable to malicious attacks that can have widespread impact. For example, a remote injection of a virus at a local station in a power network can propagate to other parts of the network, triggering responses that could lead to power outage in a bid to contain the problem. Right now, beyond isolating compromised points, there is not a holistic approach to protect network CPSs (such as the smart grid) from virus and malware attacks. The goals here are to i) model and characterize bio-defense mechanisms in networked CPS, using domain specific, realistic propagation models; and ii) propose not only techniques for securing network CPS against attacks, but also a framework for network design that increases robustness by optimally decentralizing key operations to enhance its resilience. Networks of smart devices (more popularly known as Internet of Things (IoT)) are of particular interest in this area, given their tight integration with our social and personal lives.
Uncertainty and Mechanism Design in Networks
By interacting with systems around us, humans are sometimes players and decision makers in networks, which many times have uncertainties and stochastic dynamics inherent in their operation. An example is the smart grid where ramping up the integration of renewable energy sources and storage systems require consumers’ participation in demand response programs, given the intermittent nature of renewable energy generation.
Coordinating the operation of such a system is further challenging, given the unclear and strategic actions that consumers may take at certain times. The overarching goal is to study human/agent behavior, particularly how to design operation mechanisms that yield positive social and economic outcomes, in networked population where agents are affected by the actions of others. This problem is also important in the context of network security and protection, where an agent’s lax attitude toward security or privacy in a network could be catastrophic to the entire network. This project studies design and coordination of distributed systems to achieve desired outcomes when participants are rational, strategic or sometimes strategically irrational in their decisions/actions, and affected by certain exogenous factors the operator has no control over.
Limited-Communication Optimization in Large-scale Systems
My current research involves designing limited-communication methods for distributed resource management in large systems. Motivated by problems in power distribution networks, we considered the general problem of distributed resource allocation; and using gradient-based methods, we showed how a system coordinator can limit communication in coordinating allocation of electric power to a group of heterogeneous users. Our approach considered a general class of quantized gradient methods where the gradient direction is approximated by a finite quantization set.
The method quantizes and broadcast the gradient direction as a single bit of information that encodes the coordination signal to agents from which they make local decisions on how much to consume of the shared network resource. Furthermore, we presented a convergence rate analysis that connects the fineness of the quantization with number of iterations needed to reach predefined solution accuracy. These preliminary results clearly indicate that communication in distributed coordination of large-scale systems can be made efficient.
Control of Epidemic Dynamics in Networks
In this area, we studied the problem of mitigating the spread of generic viruses in networks. We modeled infection propagation in networks using the Susceptible-Infected-Susceptible (SIS) network epidemic model and proposed a convex optimization framework for allocating control resources to secure an infected network – including directed networks with positively weighted edges of arbitrary structure.
The approach was to express the epidemic control problem in terms of spectral conditions involving the Perron-Frobenius eigenvalue of the spread dynamics. This framework enabled formulation of the epidemic control problem as a Geometric Program, for which we derived a convex characterization. Via a distributed Alternating Direction Method of Multipliers (ADMM) algorithm, we proposed a fully distributed solution to the resource allocation problem; enabling each agent to locally compute its optimum allocation of vaccines and antidotes needed to collectively, globally contain the spread of an outbreak, via local exchange of information with its neighbors.
In the past, I also looked at problems at the intersection of control and network science.
Please see my publications page for relevant, related papers.
Photo credits Disclaimer: Any photo (from an external source) used on this page is courtesy of the link associated with it.