This paper presents a novel model-based deep reinforcement learning architecture for dynamic portfolio optimization, addressing the challenge of automating wealth allocation across consecutive trading periods based on investors' return-risk profiles. The authors design an autonomous trading agent that makes periodic investment decisions based on a global objective, moving beyond purely model-free RL approaches.

The proposed architecture integrates three specialized modules: an Infused Prediction Module (IPM) that enhances decision-making capabilities, a Generative Adversarial Data Augmentation Module (DAM) that generates synthetic data to improve training, and a Behavior Cloning Module (BCM) that incorporates imitation learning techniques. The framework is compatible with both on-policy and off-policy RL algorithms and includes a real-time back-testing and execution engine that interacts with the RL agent. Using historical real financial market data with practical trading constraints, the authors demonstrate that their model-based approach achieves superior performance in terms of robustness, profitability, and risk-sensitivity compared to both baseline trading strategies and prior model-free RL agents.