A faster way to simulate economies with many decision makers
Researchers created JaxMARL-BC to train economic models in seconds instead of months.
Researchers built JaxMARL-BC to simulate economies. This tool uses multi-agent reinforcement learning (MARL) to include many different decision makers.
Multi-agent reinforcement learning (MARL) is a way to populate simulated economies with many heterogeneous decision makers. The system is JAX-accelerated. This means it trains economies in seconds. It is between 75x and 724x faster than the reference implementation.
The tool scales to 200,000 independent learning households. It works with the Multi-Agent Reinforcement Learning Business Cycle (MARL-BC). This model places deep reinforcement learning households in a real business cycle economy.
Macroeconomics has faced two main limitations. There was a lack of a shared frame of reference to connect MARL systems to dynamic stochastic general equilibrium (DSGE) models. Training times were also very long.
The researchers demonstrated value through a proof-of-concept calibration. This experiment used an extended model with heterogeneous returns on capital. It took two hours, but previously it would have taken over a month.
Key facts
- JaxMARL-BC is between 75x and 724x faster than the reference implementation.
- The system scales to 200,000 independent learning households.
- A two-hour experiment would have previously taken over a month.