Objective Human decision-making stem from the cooperation between model-free (MF) and model-based (MB) system. The MF system supports habitual behavior following the reinforcement learning (RL) rule. Instead, the MB system utilizes a “world model” to facilitate planning and rational behavior. The MB system is attributed to the prefrontal function, such as working memory (WM). However, the role of WM is not explicitly modeled in the traditional MB-MF framework, leaving the computational role of working memory ambiguous. We hypothesized that (1) WM is the buffer of value updating of the MB system; (2) WM mediates the role of prefrontal function on rational behavior. We build a new RLWM model with the WM as the value buffer of the MB system and test the model directly. Methods One-hundred and seventy-four healthy adults (young adults: 60; old adults: 114) attended the study and completed a revised two-step decision-making task. Besides, we also measured subjects’ prefrontal function using the working memory capacity test, Wisconsin Card Sorting Test, and Attention network test. Results Our results support the superiority of the RLWM model over models without the WM component. Individual analysis revealed that WM could predict subjects’ MB behavior and mediate the effect between prefrontal function and MB control. Moreover, WM protects people from the decreasing effect of aging on rational behavior.Conclusion In summary, the present study supports the WM as a value buffer for the MB system and advocates the WM as a protecting factor for aging effect on decision-making.
Keywords: Decision Making; aging; Working Memory; Model-Based; Reinforcement learning