Analysis of a Hazard Prediction Model by Means of Variable Importance

Abstract

Hazard prediction is an important element for intelligent robotic transporters to detect potential hazards like road roughness, drivability, and positive/negative obstacles from features obtained by sensor measurements. Analysis results by means of variable importance are presented for a hazard prediction model learned by random forests. Mean decrease accuracy (MDA) provides a quantitative feature importance estimation that explains which features are influential for the prediction model to make predictions. Partial dependence plot provides a qualitative explanation about how values of an important feature are used by the model. A data-driven feature selection method to find a threshold of important features by exploiting MDA is introduced. Those properties give an insight into the domain knowledge learned by the hazard prediction model as well as a reason why a prediction is returned. Explanation of inside mechanism of intelligent robotic systems is a key factor for the acceptance by societies.