Improving modeling of other agents for behavior prediction using stereotypes and compactification of observations

Abstract

This thesis presents two improvements to the Observed Situation-Action Pairs and the Nearest Neighbor Rule (OSAPs and NNR) modeling method. Reevaluative stereo­typing with switching deals with poor prediction accuracy resulting from modeling with few OSAPs by using a stereotype to model others, and incorporating periodic reevaluations and the ability to switch between stereotypes or to the basic modeling method to ensure the validity of a chosen stereotype and of the stereotyping process itself. Compactification of OSAPs through kd-tree structuring deals with poor modeling efficiency resulting from modeling with many OSAPs by structuring OSAPs according to a kd-tree and using a Pseudo-Approximate Nearest Neighbor search for modeling. Our experiments shows that using a correct stereotype improves modeling performance and reevaluations and switching prevent incorrect stereotypes from causing substantial damage, and that compactification can improve modeling efficiency - but may affect modeling performance.