محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| التنسيق: | Recurso digital |
| اللغة: | |
| منشور في: |
Zenodo
2026
|
| الموضوعات: | |
| الوصول للمادة أونلاين: | https://doi.org/10.5281/zenodo.19076073 |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
جدول المحتويات:
- <p dir="ltr">This document provides the formal theoretical foundation for the Hybrid Meta-Controller (HMC) architecture introduced in the main paper Entropy-Induced Collapse in Learned Decision Policies under Partial Observability and Architectural Mitigation. While the primary work focuses on empirical validation and system design, this supplement establishes the ontological and informational constraints governing epistemic stability in decision-making agents.</p> <p dir="ltr">We show that policy collapse under partial observability is structurally unavoidable under finite informational constraints, and cannot be eliminated by improved training procedures, optimization methods, or increased model capacity. Under a minimal set of decision-theoretic axioms, we prove the existence of a critical informational bound, denoted αcrit, below which no purely inductive policy can guarantee long-term viability. This bound arises necessarily from the interaction between finite observation channels and viability-preserving state distinctions, rather than from empirical or algorithmic limitations.</p> <p dir="ltr">We further show that asymptotic stability under partial observability requires a qualitative architectural transition, in which control shifts from inductive inference to non-inductive (structural) regulation once informational conditions fall below this bound. This necessity is formalized as the αcrit Principle of Epistemic Viability (PEV), which characterizes a fundamental constraint on the persistence of decision-making agents operating under informational limits.</p> <p dir="ltr">Together, these results establish a non-normative, information-theoretic account of epistemic stability grounded in agent viability, providing necessary structural conditions for the design of decision architectures capable of operating safely under partial observability.</p> <p dir="ltr"> </p>