Navigating Uncertainty Measures in Reinforcement Learning

Today we’re joined by Rishabh Agarwal, a research scientist at Google Brain in Montreal. In our conversation with Rishabh, we discuss his recent paper Deep Reinforcement Learning at the Edge of the Statistical Precipice, which won an outstanding paper award at the most recent NeurIPS conference. In this paper, Rishabh and his coauthors call for a change in how deep RL performance is reported on benchmarks when using only a few runs, acknowledging that typically, DeepRL algorithms are evaluated by the performance on a large suite of tasks. Using the Atari 100k benchmark, they found substantial disparities in the conclusions from point estimates alone versus statistical analysis. We explore the reception of this paper from the research community, some of the more surprising results, what incentives researchers have to implement these types of changes in self-reporting when publishing, and much more.

The complete show notes for this episode can be found at twimlai.com/go/559