Reinforcement Learning for Economic Policymaking

Deep reinforcement learning has shown promising results in games with a limited number of agents such as Chess, Go, Poker, etc. We develop methods to go beyond few agent settings and apply it to microeconomic simulations for finding general equilibria and optimal agent policies.

Several situations encountered in social and economic settings can be modeled as a game between multiple players. Multi-agent reinforcement learning (MARL) is a promising technique to simulate these situations and gain novel insights for improved policy-making and assisting individual decision-making. In this project, we develop a parallel computing framework and propose a new algorithm to scale MARL to millions of agents. Sample applications: planning of EV charging infrastructure, transportation networks, supply-chain networks, financial networks, economic networks, digital networks.