Improved Replicable Boosting with Majority-of-Majorities

Replicability of an algorithm is a property introduced as a reaction to what is called the reproducibility crisis. Multiple Nature articles have pointed out the issue of researchers not being able to replicate findings [Baker, 2016, Ball, 2023]. As a supplement to implementing better research practices in order to ensure replicability, Impagliazzo et al. [2022] introduced the concept of replicability as a property of algorithms themselves. Informally, an algorithm is replicable if it, with high probability, outputs the same result when run with different input data drawn from the same distribution.