Machine learning for automation usage prediction: identifying critical factors in driver decision-making

Abstract

Inappropriate automation usage is a common cause of incidents in semi-autonomous vehicles. Predicting and understanding the factors influencing this usage is crucial for safety. This study aims to evaluate machine learning models in predicting automation usage from behavioral data; and analyze how workload, environment, performance, and risk influence automation usage for different conditions. An existing dataset from a driving simulator study with 16 participants across four automation conditions (Speed High, Speed Low, Full High, and Full Low) was used. Five machine learning models were trained, using different splitting techniques, to predict automation usage. The input to these models were features related to workload, environment, performance, and risk, pre-processed and optimized to reduce computational time. The best-performing model was used to analyze the impact of each factor on automation usage. Random Forest models consistently demonstrated the highest prediction power, with accuracy exceeding 79% for all conditions, providing a robust foundation for enhancing vehicle safety and optimizing human-automation collaboration. Additionally, factors influencing automation usage ranked: Workload>Environment>Performance>Risk., contrasting with literature on pre-drive intentions to use automation. This study offers insights into real-time prediction of automation usage in semi-autonomous vehicles and quantifies the importance of key factors across different automation conditions. The findings reveal variations in prediction accuracy and factor importance across conditions, providing valuable implications for adaptive automated driving system design. Additionally, the hierarchy of factors influencing automation usage reveals a contrast between real-time decisions and pre-drive intentions, emphasizing the need for adaptive systems in dynamic driving conditions.