We propose a scalable method for computing global solutions of nonlinear, high-dimensional dynamic stochastic economic models. First, within a time iteration framework, we approximate economic policy functions using an adaptive, high-dimensional model representation scheme, combined with adaptive sparse grids to address the ubiquitous challenge of the curse of dimensionality. Moreover, the adaptivity within the individual component functions increases sparsity since grid points are added only where they are most needed, that is, in regions with steep gradients or at nondifferentiabilities. Second, we introduce a performant vectorization scheme for the interpolation compute kernel. Third, the algorithm is hybrid parallelized, leveraging both distributed- and shared-memory architectures. We observe significant speedups over the state-of-the-art techniques, and almost ideal strong scaling up to at least $1,000$ compute nodes of a Cray XC$50$ system at the Swiss National Supercomputing Center. Finally, to demonstrate our method's broad applicability, we compute global solutions to two variates of a high-dimensional international real business cycle model up to $300$ continuous state variables. In addition, we highlight a complementary advantage of the framework, which allows for a priori analysis of the model complexity.
翻译:我们提出一种可扩缩的方法来计算全球非线性、高维动态随机经济模型的解决方案。 首先,在一个时间迭代框架内,我们使用适应性高维模型代表方案,与适应性稀疏的电网相结合,利用适应性、高维模型代表方案,与适应性稀疏的电网相结合,应对全方位诅咒的无处不在的挑战;此外,单个元件功能的适应性会增加广度,因为电网点只在最需要的地区,即在梯度陡峭或非差异性高的地区,才增加电网点。第二,我们为内置内核计算引入一种性向导矢量的矢量计算方案。第三,算法是混合的,同时利用分布式模型和共享式模型结构。我们观察到在最新技术上的重大加速,并且几乎理想地大幅提升到至少1 000美元,计算出瑞士国家超级计算中心Cray XC$系统50美元的节点。最后,为了展示我们的方法的广泛适用性,我们将全球的矢量计算出两种高位国际实际模型的变量。第三,我们将一个前复杂度模型的连续的模型推算出一个三千位变量。