Approaching ORBIT: Examinations of Variant Agnostic Information Traversals in ML

作者：Pascal Tohouri · 发表于：Figshare · 年份：2026 · DOI：10.6084/m9.figshare.32826848.v10 · 研究领域：Computability, Logic, AI Algorithms、Constraint Satisfaction and Optimization、Explainable Artificial Intelligence (XAI)

Author's Note: These notes are early working drafts, posted to fix a public version and invite criticism. As mentioned within, each is intentionally speculative/exploratory. Several claims are currently at risk of being false or in tension with impossibility results (acknowledged where known), and the formalism is incomplete. Please read these outputs as containing conjectures and open problems, not results. Summary: The ORBIT project seeks a representation agnostic search algorithm. ML search (and instance -recoverability) friction arises from (i) an object's structure, and (ii) the object's representation during access. For example, parameter optimisation over discrete sets may become more tractable in a symmetric vector space. One endows the searcher with candidate representations that are then switched between due to local evidence and the experimentalist's utility. The most general claims (e.g., that all architectures have differently performant representations on arbitrary non-convex problems, or that instance-recoverability is always feasible) likely do not hold. So, this project examines those theoretical limits. Weaker hypotheses may nonetheless hold. For example, given an unrestricted (locally reachable) representation class and fixed compute budget, a class of ORBIT algorithms may achieve lower regret than any strictly fixed representation (in hindsight). This project also examines those claims, building on the work of Chen et al. (2015), Liu et al. (2018), and Gav...