AAAI Publications, 2017 AAAI Spring Symposium Series

Font Size: 
Learning to Predict Driver Behavior from Observation
Santiago Ontanon, Yi-Ching Lee, Sam Snodgrass, Flaura K. Winston, Avelino J. Gonzalez

Last modified: 2017-03-20


This paper focuses on modeling and predicting human driving behavior, with the long term goal of anticipating the behavior of the driver before dangerous situations occur. We formulate this problem as a Learning from Demonstration problem, and show how standard supervised learning methods do not perform well in this task. The main contribution of this paper is a new approach we call {\em indirect prediction}. The key idea of "indirect prediction" is not to predict the behavior directly, but rather to build a model that predicts how certain features of the world state will change over time, and then use those to predict the necessary behavior in order to achieve those changes. We show how this apparently counterintuitive idea directly addresses one of the key reasons for which supervised learning does not perform well for LfD. In addition, we show how using ideas from context-based reasoning can also improve the accuracy of behavior modeling.


Learning from Demonstration, Driving Behavior

Full Text: PDF