Integrating Reinforcement Learning, Response Surface Methodology, and Neural Networks into Agent-Based Modeling for Dynamic Investment Decisions

Abstract

The recent growth in data-driven and simulation-driven decision-making has attracted researchers to build intelligent and more adaptive agent-based models.  However, smarter policy modelling in complex economic systems remains challenging, with significant room for improvement. This paper develops an Artificial Neural Network (ANN) based Decision Support System (DSS) using Multiagent Reinforcement Learning (MARL) for dynamic investment decision analytics within a complex economic system. It embeds Reinforcement Learning (RL) into an Agent-Based Model (ABM) of business investments, enabling dynamic and adaptive decision-making. RL algorithms, including Q-Learning, Deep Q-Networks, and Proximal Policy Optimization (PPO), are integrated into the working behaviour of investor agents, as well as investment alternatives, to reinforce decision intelligence. A comparative evaluation of the results obtained from the four models demonstrates that the PPO-ABM yields the best output, characterized by higher investor wealth and lower annual failure risks. A formal factorial experiment with three levels of input factors is then designed in the PPO-ABM, which generated a massive amount of experimental data. Response Surface Methodology (RSM) is used to develop predictive mathematical models and response surfaces for the eight output variables of the economic system.  An ANN model is trained on the generated big data, and a recommender system is accordingly designed to facilitate intelligent decision-making and policy modelling. The developed framework can be utilized by researchers, policymakers, investors, and strategic decision-makers to conveniently test different economic scenarios and generate system-level results of the adopted policies.