A0486
Title: Multi-agent deep reinforcement learning and LLM-augmented frameworks for economic policy simulation
Authors: Tohid Atashbar - IMF (United States) [presenting]
Abstract: The application of multi-agent deep reinforcement learning (MADRL) and LLM-assisted MADRL is explored in economic policy simulation. It is argued that multi-agent modeling is the most natural approach for analyzing the complex, multi-agent nature of economies, with learning from errors mirroring real-world economic entities' trial-and-error processes. Basic concepts are introduced in MADRL, and modeling considerations are discussed. A simple framework is presented with six agents. The integration of LLM-based decision-making and AI-to-AI communication in MADRL frameworks is also examined. As a proof of concept, an LLM-augmented MADRL framework is showcased, simulating a 4-sector economy with three distinct household preferences. While the approach shows promise for enriching macroeconomic modeling tools, substantial foundational work remains in this nascent field.
