JaxMARL-BC: A Multi-Agent Reinforcement Learning Macroeconomic Simulator Trained in Seconds

This paper releases the open-source JAX-accelerated simulator JaxMARL-BC for training multi-agent reinforcement learning business cycle models, with training speeds 75 to 724 times faster than reference implementations, scalable to 200,000 independent learning households. The authors validate the simulator's correctness under known DSGE limits, characterize its scaling properties through a simple cost law, and complete a proof-of-concept calibration of an extended model with heterogeneous capital returns in a two-hour experiment, whereas the equivalent experiment previously required over a month.