Some theoretical questions about BMPO

[MeL0XIA]

This is a great work！My question is about how the assumption in the paper that the boundaries of forward and backward TVD are equal holds true. Furthermore, in many control environments, the system is not reversible. This means that multiple past states can lead to the same current state, making backward prediction undefinable. How is this problem addressed?
Looking forward to your reply！

[hanglai]

Thanks for your interest! We assume the boundaries of the forward and backward models are equally same, since in many environments (include the tasks used in the paper), given the current state and past action, the past state is unique and predictable. Note that the backward model predict s, condition on s' and a, not just s'. Hopefully this helps answer your question. Best wishes! laihang ***@***.*** 原始邮件 发件人：MeL0XIA ***@***.***> 发件时间：2026年5月10日 18:59 收件人：apexrl/bmpo ***@***.***> 抄送：Subscribed ***@***.***> 主题：[apexrl/bmpo] Some theoretical questions about BMPO (Issue #10) MeL0XIA created an issue (apexrl/bmpo#10) This is a great work！My question is about how the assumption in the paper that the boundaries of forward and backward TVD are equal holds true. Furthermore, in many control environments, the system is not reversible. This means that multiple past states can lead to the same current state, making backward prediction undefinable. How is this problem addressed? Looking forward to your reply！ — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you are subscribed to this thread.