A digital twin-based comparative reinforcement learning framework for personalized behavioral recommendation.

Area of Science:

• Digital Health
• Artificial Intelligence
• Behavioral Science

Background:

• Promoting healthy lifestyles requires adaptive systems, but data fragmentation, privacy, and ethical concerns hinder personalized recommendations.
• Current methods face challenges in dynamic behavioral and environmental condition adaptation.

Purpose of the Study:

• To propose a digital twin-driven reinforcement learning framework for personalized behavioral recommendations.
• To address challenges in developing and evaluating adaptive recommendation systems using simulated environments.

Main Methods:

• Formulated personalized behavioral recommendation as a stochastic Markov Decision Process (MDP).
• Generated synthetic longitudinal behavioral trajectories using a digital twin simulator.
• Implemented and compared various reinforcement learning (RL) paradigms (e.g., Q-learning, DQN).

Main Results:

• Richer state representations and context-dependent actions improve high-capacity RL models over simpler baselines.
• Demonstrated a reproducible method for comparing RL dynamics, performance, and computational costs.
• Validated the framework's ability to balance compliance gains with constraint violation penalties.

Conclusions:

• The digital twin-driven RL framework offers a privacy-preserving solution for personalized behavioral recommendations.
• The simulation environment allows for robust evaluation of adaptive recommendation systems.
• Future work can leverage this framework for more effective digital health interventions.