[1Win4] Poster session 1

[1Win4-104] Competitive Multi Agent Reinforcement Learning for Artificial Market Simulations

〇Ryuji Hashimoto¹, Yuri Murayama¹, Kiyoshi Izumi¹ (1.Graduate School of Engineering, The University of Tokyo)

Keywords:Multi-agent systems, Reinforcement learning, Artificial market simulation

Agent-based models (ABMs) are technologies to describe complex systems, such as financial markets, through autonomous agent interactions. Yet, traditional ABMs often rely on manually defined rules, raising concerns about the validity of their results due to subjective assumptions. Multi-agent reinforcement learning (MARL) addresses this limitation by enabling agents to learn adaptive behaviors that maximize utility. However, existing MARL-based ABMs consider only single-dimensional heterogeneity, limiting their realism. To overcome this limitation, we propose a novel approach that integrates multi-dimensional heterogeneity into MARL and develop a financial market simulation framework. Our method encodes agent characteristics—such as risk aversion, information accessibility, and time discounting—into their observations, diversifying behaviors and capturing complex market dynamics. We enhance scalability by optimizing a shared policy for agents with varied preferences. Experiments demonstrated that our approach captures the stylized facts, commonly observed phenomena in financial markets.