Apprenticeship Learning of Ship Behavior in Crowded Area by Dimension Compression

Abstract

Abstract-There is a strong demand for autonomous ship navigation systems in maritime logistics. Such systems need to be able to forecast behaviors of other ships accurately to avoid collisions. Here, time-series of ship positions, called AIS data, can be used in apprenticeship learning (AL) by defining an object map created from the data as a state and the turning direction of the ship as an action. However, when we analyzed 1 months' worth of AIS data, none of the generated path data took actions in the same state pattern twice. This paper proposes to use a Co-Moving Frame (CMF), a local segment of the environment on a small timescale. CMF improved the effectiveness of the data usage, and as a result, AL forecast paths of ships with 81.2% accuracy when applying CMF. This result is 29.2% better than that of a state transition model generated from the same dataset without applying CMF.