Albert Thomas

Research engineer in machine learning at Huawei


Fair model-based reinforcement learning comparisons with explicit and consistent update frequency

Published May 7, 2024

This is a blog post that we wrote with my colleagues Abdelhakim Benechehab, Giuseppe Paolo, and Balázs Kégl for the Third Blog Post Track at ICLR 2024. We discuss the impact of implicit update frequencies on model-based reinforcement learning benchmarks. These frequencies can introduce ambiguity, making it difficult to evaluate algorithms effectively. While optimizing the update frequency can sometimes improve performance, real-world applications often impose constraints that only allow updates between system deployments. Our blog emphasizes the need for consistent update frequencies across different algorithms to provide clearer comparisons under realistic conditions.

You can check out the blog post here.