Adrian Kulesza

Working Papers

Combinatorial Discrete Choice with Deep Reinforcement Learning
Adrian Kulesza
November 2024
Code coming soon.
Abstract: I utilize new computational methods to study how economic agents solve combinatorial optimization (CO) problems, where an optimal solution is selected from a large, discrete set. In structural trade/spatial models CO problems are ubiquitous from firms choosing where to source inputs and open plants to social planners choosing where to allocate infrastructure and enact policies. Normally, these problems are approached using either heuristics or specialized algorithms informed by model assumptions and precise parameterizations.CO problems are ubiquitous in structural trade and spatial models, from firms deciding where to source inputs and open plants to social planners determining where to allocate infrastructure and enact policies. I use a machine learning model to approximate policy functions that learn to solve CO problems through repeated interaction with a simulated economic environment. I benchmark this approach to existing algorithms for several CO problems from trade, often yielding either optimal or superior policies with competitive computational times. I then demonstrate how this method can be applied to models with rich interdependencies, for which current methods do not work. I estimate a model of export market entry with complementarity in fixed costs and substitutability through increasing marginal costs.