Agent Decision Model Based on Deep Reinforcement Learning in Artificial Stock Market

In order to explore the impact of reinforcement learning on agent decision in agent-based computational finance models, the dueling network architectures is introduced to construct the stock market agent decision-making model based on the deep Q network, and the stock market trading strategy is learned by maximizing the return on investment. The realistic trading data are used for model training. The experiment shows that agent can determine when to buy and sell to obtain positive rewards, and can learn the investment strategy of holding for a long time to increase the final return. The research in this paper enables the agent-based computational finance model to simulate the microstructure of the stock market in a more realistic way.