本研究提出了一種基於強化學習和多代理人架構的動態資產配置方法。在建構典型的股債資產配置投資組合後，採用多代理人架構，由資金管理代理人控管投資組合的資產配置，並將資金交由股票交易代理人及債券交易代理人分別進行股票ETF及債券ETF之交易。此架構促進專業分工，使各代理人能專注其特定任務，提升學習效率和決策品質。研究選用適用於連續動作空間的DDPG演算法實施細膩且精準的動態資產配置。;This study proposes a dynamic asset allocation method based on reinforcement learning and a multi-agent framework. After constructing a typical stock-bond asset allocation portfolio, a multi-agent framework is adopted, where a fund management agent controls the asset allocation of the portfolio, and then allocates funds to stock trading agent and bond trading agent to trade stock ETFs and bond ETFs, respectively. This framework promotes professional specialization, allowing each agent to focus on their specific tasks, thereby enhancing learning efficiency and decision-making quality. The DDPG algorithm, suitable for continuous action spaces, is selected to implement fine and precise dynamic asset allocation.
