JAX-Accelerated Multi-Agent Macro Model Trains in Seconds
ORIGINAL / JaxMARL-BC: Multi-Agent Reinforcement Learning for Macroeconomic Modelling, in Seconds
New research rewrites the MARL-BC macroeconomic simulator in JAX, achieving 75–724x faster training and scaling to 200,000 independent learning households, cutting calibration experiments from over a month to two hours.
01 ABSTRACT
JaxMARL-BC is a JAX-accelerated simulator for training multi-agent reinforcement learning macroeconomic models. The authors report training speeds 75–724 times faster than the reference implementation, scaling to 200,000 independent learning households, and completing a calibration experiment in two hours that previously took over a month. Correctness is verified against known DSGE limits, and a simple cost scaling law is characterized.
02 KEY FINDINGS
- Training is 75–724x faster than the reference implementation and scales to 200,000 independent learning households.
- The model reduces to classic DSGE models in specific limits, which is used to verify correctness.
- A proof-of-concept calibration of an extended model with heterogeneous capital returns cuts runtime from over a month to two hours.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG · ARXIV Q-FIN