What AI Can Do in Economics?
Paper Session
Tuesday, Jan. 5, 2021 10:00 AM - 12:00 PM (EST)
- Chair: Paul Milgrom, Stanford University
Optimal Targeted Lockdowns in a Multi-Group SIR Model
Abstract
We study targeted lockdowns in a multi-group SIR model where infection, hospitalization and fatality rates vary between groups—in particular between the “young”, “the middle-aged” and the “old”. Our model enables a tractable quantitative analysis of optimal policy. For baseline parameter values for the COVID-19 pandemic applied to the US, we find that optimal policies differentially targeting risk/age groups significantly outperform optimal uniform policies and most of the gains can be realized by having stricter lockdown policies on the oldest group. Intuitively, a strict and long lockdown for the most vulnerable group both reduces infections and enables less strict lockdowns for the lower-risk groups. We also study the impacts of group distancing, testing and contract tracing, the matching technology and the expected arrival time of a vaccine on optimal policies. Overall, targeted policies that are combined with measures that reduce interactions between groups and increase testing and isolation of the infected can minimize both economic losses and deaths in our model.Will Artificial Intelligence Replace Computational Economists Any Time Soon?
Abstract
Artificial intelligence (AI) has impressive applications in many fields (speech recognition, computer vision, etc.). This paper demonstrates that AI can be also used to analyze complex and high-dimensional dynamic economic models. We show how to convert three fundamental objects of economic dynamics -- lifetime reward, Bellman equation and Euler equation -- into objective functions suitable for deep learning (DL). We introduce all-in-one integration technique that makes the stochastic gradient unbiased for the constructed objective functions. We show how to use neural networks to deal with multicollinearity and perform model reduction in Krusell and Smith's (1998) model in which decision functions depend on thousands of state variables (we literally feed distributions into neural networks!) In our examples, the DL method is reliable, accurate and linearly scalable. Our ubiquitous Python code, built with Dolo and Google TensorFlow platforms, is designed to accommodate a variety of models and applications.Solving High-Dimensional Dynamic Programming Problems using Deep Learning
Abstract
To answer a wide range of important economic questions, researchers must solve high-dimensional dynamic programming problems. This is particularly true in models designed to account for granular data. To break the ``curse of dimensionality'' associated with these high-dimensional dynamic programming problems, we propose a deep-learning algorithm that efficiently computes a global solution to this class of problems. Importantly, our method does not rely on integral approximation, instead efficiently calculating exact derivatives. We evaluate our methodology in a standard neoclassical growth model and then demonstrate its power in two applications: a multi-location model featuring 50 continuous state variables and a highly nonlinear migration model with 75 continuous state variables.Discussant(s)
Guillaume Haeringer
,
Baruch College
Kurt Mitman
,
Stockholm School of Economics
Per Krusell
,
Stockholm School of Economics
SeHyoun Ahn
,
Norges Bank
JEL Classifications
- C6 - Mathematical Methods; Programming Models; Mathematical and Simulation Modeling
- D0 - General