Collaborative Research: Adaptive Gaussian Markov Random Fields for Large-scale Discrete Optimization via Simulation

Major federal agencies, including the Department of Veterans Affairs, Department of Defense, Department of Homeland Security, Federal Aviation Administration, Department of the Treasury, Internal Revenue Service, Centers for Medicare and Medicaid Services, Department of Health and Human Services, and others, seek government and non-government assistance with the application of scientific, data-driven methods to help them execute effectively on their critical missions. Because their mandate is typically large-scale, complex, and involves inherent uncertainty, computer simulation is often the only tool for representing their problems in a comprehensive way. Similar problems occur in the private sector, especially in health care delivery, computer networks, warehousing and distribution, and transportation systems. Unfortunately, "large-scale, complex, and involving inherent uncertainty" are the features that make "optimizing" a simulated system hard, particularly when the decisions are how to allocate discrete units of resources such as personnel, vehicles and facilities. The proposed research marries high-performance computing, smart numerical methods, and state-of-the-art statistical methodology to significantly increase the size and complexity of simulated systems that can be optimized. As a result, agencies such as those listed above will be able to more fully solve their "system of systems" resource-allocation problems using computer simulation. The proposed research tackles statistical and computational challenges that arise in solving large-scale stochastic optimization problems when the objective function may only be evaluated by executing a stochastic simulation. Such optimization problems are often with respect to a high-dimensional, discrete-valued decision variable in a large solution space. The modeling flexibility of simulation comes at a cost: arbitrarily complex stochastic simulations may not be optimized using tools from mathematical programming. As a result, the scale of problems that can currently be solved by simulation with an optimality gap guarantee is limited. The investigators propose to create theory, algorithms and software for large-scale discrete-decision-variable simulation optimization that converge to the global optimum asymptotically, and provide optimality-gap inference when terminated. The proposed methods are based on inferential optimization, which models the unknown objective function by a Gaussian Markov Random Field (GMRF), a type of Gaussian Process defined by a graph on the discrete solution space; the investigators have shown that GMRFs provide better inference for a discrete problems than Gaussian processes defined on a continuous domain. The conditional distribution of a GMRF provides inference for selecting solutions to simulate and for search termination when the inferred optimality gap is small. However, the computational cost of numerical linear algebra increases faster than the number of feasible solutions. To facilitate the solution of large-scale problems, three core topics are proposed: exploiting high-performance computing; creating a restricted search scheme and tailored computational linear algebra that significantly reduces the computations in GMRF updates; and attacking limits on dimensionality via an adaptive multi-resolution GMRF and projections to lower dimensions. This award will provide support of graduate student training through research. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.