Learning Sparse Evidence-Driven Interpretation to Understand Deep Reinforcement Learning Agents

Recent advances in machine learning require interpretability and explainability for reliable and trustworthy systems. However, explanations of machine learning models are often hard to achieve given the large amount of information from the complex machine learning models. Evidence-driven reinforcement learning provides snapshot images to understand the learning experiences and the learned behaviors; however, it requires human labor to analyze a large number of retrieved snapshot images. Imposing sparsity of the evidence collection process for interpretation is, thus, significant to make human interpretation easy. In this paper, we proposed novel sparse evidence collection methods to discarding less important images for interpretation. We discuss the trade-offs between the sparsity and re-approximation accuracy and the quality of evidence in different Atari game environments.