Econ Papers Daily / 期刊论文
序贯内生网格法
The Sequential Endogenous Grid Method
摘要
Heterogeneous-agent models with multiple decisions face a computational dilemma. Joint optimization over all choices is slow; strong separability restrictions speed computation but rule out economically interesting interactions. I introduce Sequential EGM (EGM n ), which eases this tension by decomposing the problem into sequential stages so that each stage can exploit the Endogenous Grid Method’s closed-form inversion, chaining speed gains across decisions wherever separable utility or an invertible transition permits it. The resulting curvilinear grids require interpolation that respects their structure, so I develop ENGINE (ENdogenous Grid INterpolation and Extrapolation), which exploits the index structure inherited from exogenous grids, with no construction step and no triangulation. In the canonical liquid-illiquid household block, EGM n solves an order of magnitude faster than the nested method in serial and 3.7 to 7.9 times faster at sixteen threads. Replacing existing curvilinear interpolation with ENGINE cuts full-solve time 2.8- to 4.5-fold at matched accuracy.
中文摘要
具有多重决策的异质性主体模型面临计算困境。对所有选择进行联合优化速度缓慢；强可分性限制虽能加速计算，却排除了具有经济意义的交互作用。本文引入序贯EGM（EGM n），通过将问题分解为序贯阶段来缓解这一张力，使每个阶段都能利用内生网格法的闭式反演，在可分效用或可逆转移允许之处将速度收益跨决策串联。由此产生的曲线网格需要尊重其结构的插值方法，因此我开发了ENGINE（内生网格插值与外推），该方法利用继承自外生网格的索引结构，无需构建步骤，也无需三角剖分。在典型的流动性—非流动性家庭模块中，EGM n在串行计算下比嵌套方法快一个数量级，在十六线程下快3.7至7.9倍。以ENGINE替代现有曲线插值，在匹配精度下将完整求解时间缩短2.8至4.5倍。