Unfortunately I notice many RL researchers don't play or even look at the games. Many of them just look at plots of reward graphs and don't even know what things the agent has learned... You can see this in the results section of a few recent NeurIPS papers.
— hardmaru (@hardmaru) January 11, 2019
Or any other dataset, really. I bet many CV researchers have never looked through the training data images either. Same for analyzing the model errors. I think it’s one of the most valuable things you can do, but many only look at the final accuracy score.
— Denny Britz (@dennybritz) January 11, 2019
