Reinforcement Learning Induced Non-Neutrality of Monetary Policy in Computational Economic Simulation

In a Real Business Cycle model, monetary shock does not affect real variables, and economic agents are assumed to understand the model’s structure. This article shows how it is possible to build a macroeconomic agent-based simulation from standard textbook Real Business Cycle model and how to utilize reinforcement learning to drive agents’ decision making. The reinforcement learning algorithm of choice in this article is Q-learning, extended with fuzzy approximation. Q-learning is a simple algorithm based on incremental updates of estimated future rewards. As such, it circumvents introducing black boxes into the simulation and does not require strong assumptions on economic agents’ rationality and expectations. This simulation falls into Real Business Cycle model category, but the reinforcement learning driven decision making mechanism of economic agents causes monetary policy to be non- neutral in the short run