CDS&E-MSS/Collaborative Research: Sequential Design for Stochastic Control: Active Learning of Optimal Policies

Virginia Polytechnic Institute And State University, Blacksburg VA

Investigators

Abstract

This research project aims to build new cross-disciplinary algorithms that blend concepts from applied probability, control, and statistical modeling to tackle computational challenges in large-scale optimization. Creation of such new links is another step in building a next-generation of high-performance algorithms needed to meet the increasingly complex problems arising in applications as diverse as finance, energy storage and security, and the management of epidemics. The project's research agenda is grounded in two concrete application areas where it is crucial to tackle industrial-grade high-fidelity models. One is the efficient management of cycled commodity assets, including gas storage, battery storage, or fleets of power plants as energy infrastructure is transitioned to the "smart grid." A second is timely and effective response to unfolding infectious disease outbreaks, notably influenza. Both present major cross-disciplinary challenges. We see vast potential for algorithms which expand capabilities for aspects of quantitative control, and thus provide higher quality information to decision makers. Our goal is to produce a smarter, more targeted, use of random numbers in a new wave of lean stochastic solvers, and subsequently an expansion of the size of problems that can be tackled with existing computing capabilities. The educational core of the project contributes to inter-disciplinary training in mathematical sciences across undergraduate, graduate and postdoctoral levels. The collaborative initiatives will also enhance the research infrastructure through exchange of ideas between the two campuses (Univesrity of California-Santa Barbara and University of Chicago) and communities of statisticians, operations researchers and engineers. All algorithms would be documented and publicly released to the wider scientific community. Deployment of simulation based schemes remains key for control of stochastic systems that require realistic high-fidelity representations. This project will develop new Monte Carlo algorithms for a class of stochastic control problems by erecting novel bridges between dynamic control and methods of sequential design and statistical learning. Our research agenda hinges on sequential, active learning of optimal action sets, so that the algorithms adaptively allocate computing resources to better enhance fidelity of the approximated control strategies. Such targeted use of Monte Carlo simulations links approximate dynamic programming with response surface modeling, marrying two so-far disparate areas of applied mathematics and statistics. The resulting adaptive schemes will facilitate orders of magnitude savings in simulation budgets, expanding the frontier for predictive modeling and decision making under uncertainty. The proposed research will advance the theory of algorithms for dynamic control over massive multi-dimensional state spaces, where curses of dimensionality are unavoidable. Simultaneously, integration of the statistical and computational theories in this direction will open new lines of interdisciplinary quantitative research. Through enhancing knowledge discovery in large-scale control settings, the projects will facilitate transition to practice in novel contexts. With the aim of reaching out to diverse users from the mathematical, biological, physical and engineering sciences, producing general purpose open-source software via R packages is a primary deliverable of the project, and will be supplemented by a database of case studies.