Deep Learning Algorithms for Solving Finite-Horizon Models

We develop and test deep learning algorithms for solving finite-horizon models. We show how the use of neural networks and simulation help tame the curse of dimensionality. We succeed in accurately solving a consumption-saving model with 8 durable goods, convex adjustment costs and irreversible investment in a few hours on a single GPU. We verify the stability and robustness of the solution methods across multiple models with limited hyperparameter tun-
ing. We also show how one of the algorithms can be used to solve a non-convex model with both discrete and continuous choices. Our algorithms are implemented as a user-friendly open source Python package, where it is only necessary to specify high-level information such as the distribution of initial states and shocks, the transition function and the reward function.

With Jacob Røpke

Link to draft: https://drive.google.com/file/d/1txCSkwSXSQo1zl576g7MdOc2ygoK7ucK/view