Using Planning Landmarks to Control Camera Movement in DOTA 2 Games

This thesis introduces a new method for automatically finding important events from video game replays. We use these events to control the movement of the camera in the Dota 2 game. This method is based on the idea of finding landmarks, which are events that always takes place in a game. Our method automatically highlights important events in the Dota 2 game. The method combines two algorithms: one assigning weights to events and another one computing similarities between events. We discuss the motivation and implementation of these two algorithms and we test them in a transportation domain and in the Dota 2 game. Several experiments are performed in both domains and the results show that the method extracts landmarks effectively. Based on the experiments on Amazon mechanical turk, camera control based on landmarks shows potential benefits.