Presentation

Poster
:
EAD01 - Deep Equilibrium Nets
Event Type
Poster
TimeThursday, 13 June 201919:50 - 21:50
LocationHG EO Nord
DescriptionWe propose a grid-free, global solution method to compute approximate recursive equilibria for stochastic economies with a high-dimensional state space in discrete time. We use deep neural nets to approximate the equilibrium demand functions. The key innovation of our paper is to use the implied error in the optimality conditions as loss function. Consequently, training data can be generated at virtually zero cost, since neither sets of nonlinear equations nor optimization problems need to be solved. To illustrate the applicability of the proposed method, we solve for an approximate equilibrium in an overlapping generations model with 60 generations, aggregate uncertainty, and occasionally binding constraints. In our numerical experiments on a Cray XC50 "Piz Daint" system at the Swiss National Supercomputer Centre (CSCS), we obtain average relative errors in the Euler equations of the order ~ 10^(-3) when applying a densely connected deep neural net with two hidden layers.