[R] Split the Differences, Pool the Rest: Provably Efficient Multi-Objective Imitation

TLDR: The question we answer: how do you learn from experts with different objectives? Pooling all their data can lose their trade-offs; learning from each expert separately misses opportunities to share data. MA-BC pools demonstrations where observed actions don’t disagree, with upper and lower bounds on sample complexity.
Authors: Ziyad Sheebaelhamd, Luca Viano, Volkan Cevher, Claire Vernade

Arxiv: https://arxiv.org/abs/2605.12000
Github: https://github.com/ziyadsheeba/mabc

https://preview.redd.it/i20adc3z04uh1.png?width=2532&format=png&auto=webp&s=0ea80d746ea8075ca4b12203ccdcab6c2245890e