Scalable high-dimensional dynamic stochastic economic modeling

Abstract

We present a highly parallelizable and flexible computational method to solve high-dimensional stochastic dynamic economic models. Solving such models often requires the use of iterative methods, like time iteration or dynamic programming. By exploiting the generic iterative structure of this broad class of economic problems, we propose a parallelization scheme that favors hybrid massively parallel computer architectures. Within a parallel nonlinear time iteration framework, we interpolate policy functions partially on GPUs using an adaptive sparse grid algorithm with piecewise linear hierarchical basis functions. GPUs accelerate this part of the computation one order of magnitude thus reducing overall computation time by 50%. The developments in this paper include the use of a fully adaptive sparse grid algorithm and the use of a mixed MPI-Intel TBB-CUDA/Thrust implementation to improve the interprocess communication strategy on massively parallel architectures. Numerical experiments on "Piz Daint" (Cray XC30) at the Swiss National Supercomputing Centre show that high-dimensional international real business cycle models can be efficiently solved in parallel. To our knowledge, this performance on a massively parallel petascale architecture for such nonlinear high-dimensional economic models has not been possible prior to present work. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. AbstractWe present a highly parallelizable and flexible computational method to solve highdimensional stochastic dynamic economic models. Solving such models often requires the use of iterative methods, like time iteration or dynamic programming. By exploiting the generic iterative structure of this broad class of economic problems, we propose a parallelization scheme that favors hybrid massively parallel computer architectures.Within a parallel nonlinear time iteration framework, we interpolate policy functions partially on GPUs using an adaptive sparse grid algorithm with piecewise linear hierarchical basis functions. GPUs accelerate this part of the computation one order of magnitude thus reducing overall computation time by 50%. The developments in this paper include the use of a fully adaptive sparse grid algorithm and the use of a mixed MPI-Intel TBB-CUDA/Thrust implementation to improve the interprocess communication strategy on massively parallel architectures. Numerical experiments on "Piz Daint" (Cray XC30) at the Swiss National Supercomputing Centre show that highdimensional international real business cycle models can be efficiently solved in parallel. To our knowledge, this performance on a massively ...

Machine Learning for Dynamic Incentive Problems

“…Instead, this training input (and the corresponding, nonsensical training target) can be discarded-that is to say, it is not added to the training set. This is in stark contrast to grid-based methods such as "Smolyak" (see, e.g., Krueger and Kubler, (2004) and Judd et al, (2014)) or "adaptive sparse grids" (see, e.g., Brumm and Scheidegger, (2017) and Brumm et al, (2015)), where the construction of the surrogate breaks down if not every optimization problem required by the algorithm can be solved. Second, computing solutions solely on a domain of relevancethat is, W (θ), allows one to carry out VFI on complex, high-dimensional geometries without suffering from massive inefficiencies, as the computational resources are concentrated where needed.…”

Self-Justified Equilibria: Existence and Computation

“…To solve for self-justied equlibria in general, we need to repeatedly approximate and interpolate multi-variate policy function on irregularly-shapedthat is, non-hypercubic domains. In such environments, standard grid-based methods such as Smolyak (see, e.g., Krueger and Kubler (2004) and Judd et al (2014)) or adaptive sparse grids (see, e.g., Brumm and Scheidegger (2017) and Brumm et al (2015)), will fail. To this end, we will follow closely Scheidegger and Bilionis (2017) and use Gaussian process regression (GPR) (see, e.g., Rasmussen and Williams (2005) and Sec.…”

Using Adaptive Sparse Grids to Solve High-Dimensional Dynamic Models

“…Combining distributed memory parallelism with the on-node shared memory parallelism and the usage of GPUs, we are able to scale our code to at least 2,048 compute nodes with more than 60% efficiency, resulting in an overall speedup of more than 4 orders of magnitude compared to running this numerical experiment on a single CPU. More details regarding the optimization of our code and the parallel implementation can be found in Appendix D and in our companion paper (Brumm, Mikushin, Scheidegger, and Schenk (2015)).…”