A-BERF: Action-Weighted Ensemble by Bootstrapping Extremely Randomized Forest for Pre-Crash Moral Decision-Making in Autonomous Driving

Abstract

This study proposes a novel and high-precision decision-making methodology of an Action-weighted ensemble by Bootstrapping Extremely Randomized Forest (A-BERF) for the moral collision dilemma during urban autonomous driving. By simulating the pedestrian-crossing situation, the decision result from the experiment participants and the features were combined into the dataset. The performance between the tree or forest-based ensemble baseline methods and A-BERF was compared. As a result of the experiment, within the same method, the higher the dimension of the tree and the similar consideration of the unbalanced ratio of data and the weight of class, the higher the accuracy. In addition, A-BERF had the highest classification accuracy and the lowest feature bias compared to other ensemble methods using various types of datasets in validation. In addition, we confirmed that the operation time was improved compared to the random forest.