| Summary: | 碩士 === 國立交通大學 === 多媒體工程研究所 === 102 === This thesis is about the analysis and imitation of human and non-human players’ driving styles in the TORCS platform. We classify existing controller architectures as low-level controllers or high-level controllers, and we achieve better results of imitation by integrating these types of controllers.
For evaluating the imitation results, we propose a method to estimate the similarity of driving styles. It is also useful for a trained controller to adapt to a new track. We use multi-objective optimization algorithm as a method for controllers’ self-adaptation. According to our experiments, if a NPC with almost optimal speed performance is selected as the imitation target, the trained controller is able to produce similar trajectories and is only slower by 4%~14% than the imitation target in very difficult and unfamiliar tracks.
In this thesis, we also discuss whether different human players have unique driving styles. By using our proposed method for estimating driving style similarities, it is possible to observe the differences of driving styles and we can even recognize a player by analyzing the driving data. Besides, according to some proposed researches, if human players are selected as the imitation targets, the trained controllers usually crash on the tracks frequently. Therefore, the speed performance is quite low. For dealing with this issue, we propose a method to train a controller with robust driving style. Such a trained controller is able to show some driving behaviors of the target human player and seldom crashes.
|
|---|
