Understanding Human Decision Making in an Interactive Landslide Simulator Tool via Reinforcement Learning.

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Risk Management

Background:

• Human decision-making in landslide risk scenarios has been studied using the Interactive Landslide Simulator (ILS).
• Previous research indicates that feedback on landslide damages improves human decision-making.
• The application of learning theories, specifically reinforcement learning (RL), to understand these decisions remains underexplored.

Purpose of the Study:

• To model human decisions within the ILS tool using computational reinforcement learning approaches.
• To investigate the underlying mechanisms of human decision-making when faced with landslide risks.
• To evaluate the efficacy of different RL models in capturing observed human behavior.

Main Methods:

• Four distinct RL models were developed: Expectancy-Valence (EV), Prospect-Valence-Learning (PVL), PVL-2, and an EV-PU combination model.
• Model parameters were calibrated using human decision data from experiments with two distinct conditions in the ILS.
• The best-performing models were then generalized to data from a new experimental condition.

Main Results:

• The Prospect-Valence-Learning-2 (PVL-2) model demonstrated superior performance in generalizing to the new ILS condition compared to other RL models and a random model.
• Calibrated parameters from damage-feedback conditions in PVL-2 showed the most accurate predictions.
• The study provides evidence for the utility of specific RL models in explaining human choices under risk.

Conclusions:

• Reinforcement learning, particularly the PVL-2 model, offers a robust framework for understanding human decision-making in landslide risk contexts.
• Findings have implications for designing more effective risk communication and mitigation strategies.
• Computational modeling can enhance our understanding of cognitive processes in natural hazard management.