AMS Weekly Seminar | Jesus Fernandez-Villaverde

Location: Shaffer 3

When: April 9th at 1:30 p.m.

Title: The Blessings of Overparameterization

Abstract: We study the behavior of overparameterized neural networks in solving economic dynamic programming problems. Using two canonical examples (a McCall job search model and a Linear-Quadratic Regulator applied to a firm investment problem with adjustment costs), we show that overparameterization yields three key benefits: Bellman and Euler residuals decline monotonically with the number of parameters; approximate solutions converge to closed-form solutions with relative errors below 0.001; and optimization becomes increasingly stable across random initializations. These results contrast with standard regression problems, where test loss plateaus while train loss continues to fall. The self-referential structure of Bellman equations appears to interact favorably with overparameterization in ways absent from standard supervised learning. Results are robust across multiple function approximation families, including neural networks with alternative activation functions, Random Fourier Features, and Chebyshev polynomials.

Zoom link: https://wse.zoom.us/j/92366532431