• Classified by Topic • Classified by Publication Type • Sorted by Date • Sorted by First Author Last Name • Classified by Funding Source •
Learning Optimal Advantage from Preferences and Mistaking it for Reward.
W. Bradley
Knox, Stephane Hatgis-Kessell, Sigurdur Orn Adalgeirsson, Serena Booth, Anca Dragan, Peter
Stone, and Scott Niekum.
In The 38th Annual AAAI Conference on Artificial
Intelligence (AAAI), February 2024.
We consider algorithms for learning reward functions from human preferences overpairs of trajectory segments, as used in reinforcement learning from humanfeedback (RLHF). Most recent work assumes that human preferences are generatedbased only upon the reward accrued within those segments, or their partialreturn. Recent work casts doubt on the validity of this assumption, proposing analternative preference model based upon regret. We investigate the consequencesof assuming preferences are based upon partial return when they actually arisefrom regret. We argue that the learned function is an approximation of theoptimal advantage function, not a reward function. We find that if a specificpitfall is addressed, this incorrect assumption is not particularly harmful,resulting in a highly shaped reward function. Nonetheless, this incorrect usageof the approximation of the optimal advantage function is less desirable than theappropriate and simpler approach of greedy maximization of it. From theperspective of the regret preference model, we also provide a clearerinterpretation of fine tuning contemporary large language models with RLHF. Thispaper overall provides insight regarding why learning under the partial returnpreference model tends to work so well in practice, despite it conforming poorlyto how humans give preferences.
@InProceedings{brad_knox_AAAI2024,
author = {W. Bradley Knox and Stephane Hatgis-Kessell and Sigurdur Orn Adalgeirsson and Serena Booth and Anca Dragan and Peter Stone and Scott Niekum},
title = {Learning Optimal Advantage from Preferences and Mistaking it for Reward},
booktitle = {The 38th Annual AAAI Conference on Artificial Intelligence (AAAI)},
year = {2024},
month = {February},
location = {Vancouver, Canada},
abstract = {We consider algorithms for learning reward functions from human preferences over
pairs of trajectory segments, as used in reinforcement learning from human
feedback (RLHF). Most recent work assumes that human preferences are generated
based only upon the reward accrued within those segments, or their partial
return. Recent work casts doubt on the validity of this assumption, proposing an
alternative preference model based upon regret. We investigate the consequences
of assuming preferences are based upon partial return when they actually arise
from regret. We argue that the learned function is an approximation of the
optimal advantage function, not a reward function. We find that if a specific
pitfall is addressed, this incorrect assumption is not particularly harmful,
resulting in a highly shaped reward function. Nonetheless, this incorrect usage
of the approximation of the optimal advantage function is less desirable than the
appropriate and simpler approach of greedy maximization of it. From the
perspective of the regret preference model, we also provide a clearer
interpretation of fine tuning contemporary large language models with RLHF. This
paper overall provides insight regarding why learning under the partial return
preference model tends to work so well in practice, despite it conforming poorly
to how humans give preferences.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Wed Apr 17, 2024 18:42:49