Visual Analysis of DDPG Models by Exploring the Space of Experience.

Deep Reinforcement Learning (DRL) has been remarkably successful, but the lack of RL expertise and the complexity of DRLs hinder model understanding. In this paper, we focus on visual analysis of experience data to improve the interpretability of DRLs, which involves step aggregation, high-dimensional state data analysis, and spatio-temporal modeling of experience data. In addition, we introduce DDPGVis, a visual system which combines multiple views to show statics and allows users to explore the experience space, and its effectiveness is confirmed by case studies.