Cognitively-plausible reinforcement learning in epidemiological agent-based simulations

Area of Science:

• Computational epidemiology
• Cognitive science
• Public health modeling

Background:

• Human behavior significantly impacts infectious disease transmission and public health intervention effectiveness.
• Traditional agent-based models (ABMs) often oversimplify behavioral dynamics, limiting their accuracy.
• Integrating complex decision-making into simulations is crucial for realistic epidemiological modeling.

Purpose of the Study:

• To develop a novel framework for agent-based models (ABMs) that incorporates cognitively plausible reinforcement learning (RL).
• To enable dynamic behavioral adaptation in simulations without extensive data training.
• To enhance the accuracy and interpretability of epidemiological simulations.

Main Methods:

• Proposed a framework combining Adaptive Control of Thought-Rational (ACT-R) and Instance-Based Learning (IBL) for nonparametric RL in ABMs.
• Modeled mask-wearing behavior during the COVID-19 pandemic to demonstrate the framework's utility.
• Analyzed the influence of local versus global social cues on behavior and disease transmission.

Main Results:

• Local social cues strongly correlated with clustered mask-wearing behavior (slope = 0.54, r = 0.76).
• Reliance on global cues alone resulted in weakly disassortative patterns (slope = 0.05, r = 0.09).
• Demonstrated the framework's scalability and cognitive interpretability in simulations.

Conclusions:

• The novel framework effectively integrates adaptive decision-making into epidemiological simulations.
• Local information plays a critical role in coordinating public health compliance.
• The approach offers actionable insights for public health policy and intervention strategies.