2:00 PM - 2:20 PM
[4D2-OS-18c-01] Filtering Environmental Information for Automatic Driving in Urban Area
Keywords:Automatic Driving, Deep reinforcement learning
This paper focuses on lack of explainability about the actions made by Deep reinforcement learning (DRL). DRL shows its superiority on tasks with multi-dimensional visual-input such as playing Atari games and navigation of robot. Also, DRL has been expected as an useful approach to realize a self-driving car. Though a lot of inputs would be available within urban environment, we could not specify the essential inputs to decide the appropriate action. The results derived from DRL are the tacit knowledge that is difficult to transfer to another system by means of writing it down as a symbolic way. Thus, human designer of self-driving operation may be reluctant to accept and utilize these results.
For the above reason, as a first stage to reach a symbolic representation, we propose a filtering method to specify the necessary inputs for the right actions, and show the effectiveness of it via some experiments.
For the above reason, as a first stage to reach a symbolic representation, we propose a filtering method to specify the necessary inputs for the right actions, and show the effectiveness of it via some experiments.