Amber Srivastava

Research

    My research integrates ideas from three core areas — optimization, learning and control — and provides new theoretical frameworks and efficient algorithms to address a broad range of network science and planning problems. These problems span diverse disciplines and applications such as facility location with path optimization (FLPO), vehicle routing, sensor network design, manufacturing process parameter optimization, robotics, last mile delivery, job-shop scheduling, and data aggregation, classification, and clustering algorithms. The goal of my research has been to develop tools that are typically rooted in abstract formalism, and therefore address a general class of problems that span many areas, while at the same time are flexible enough to accommodate variations of specific applications.

    Cyber-Physical Manufacturing Systems - Industry 4.0

    The Industry 4.0 seeks to build efficient and flexible manufacturing capabilities that exploits the interactions between the manufacturing resources (mechanical and electrical equipments) and high-performance computing resources. The decision-making in these systems rely heavily on cyber-physical infrastructure that collects, curates, stores, and analyses data to make suitable planning and scheduling decisions, and pre-empt system failures and deadlocks. To this end, designing data-driven decision-making algorithms that are flexible to incorporating constraints and customer demands, and resilient to risks is of paramount importance. The work in this project will target various optimization problems in cyber-physical manufacturing systems and will be dealt within a much more generalized settings.

    Learning, Optimization and Control in Large Scale Networks

    Many complex systems either fall directly into the category of large scale networks, or are modelled as networks whose nodes and edges correspond to the decision variables/parameters of the original problem. Application areas such as supply chain networks, industrial process monitoring, power grids, unit commitment problem, battlefield surveillance, and software defined networks fall into this category. The problems underlying these applications belong to a subset of modeling, identification, learning, optimization, or control of these systems. Inherent to these systems are the (possibly) large scale models of the underlying network; and thus require tools and methodologies that exploit their inherent characteristics (such as sparsity or decentralized control design) to address the above problems efficiently.

    Statistical Physics and Combinatorial Optimization

    Maximum Entropy Principle (MEP), from the statistical physics literature, is often used to address combinatorial optimization problems. A peculiar characteristic of these MEP-based methods is the phase transition phenomenon. In particular, there are specific instances in the annealing procedure of these MEP-based frameworks at which the solution undergoes significant changes. We explore the utility of these phase transitions in determining certain design hyperparamters in optimization problems; for instance, estimating the true number of clusters in a data set, appropriate choice for the number of superstates in the Aggregation of Markov chains, determining the sparsity level in sparse linear regression problems, or the size (i.e., number of layers and neurons) in an Artificial Neural Network.