Solomonoff induction based on reflective oracles
Solomonoff Induction with Reflective Oracles
Translation in progress — untranslated text is omitted.
Quick read
Based on the paper PDF (AI-generated); refer to the original
A variant of Solomonoff induction based on reflective oracles, whose optimal predictor can outperform computable prediction strategies that specify only part of the data-generating process. The argument relies on the assumption that hypotheses computable from the predictor's own conditional probabilities remain within the reflective hypothesis class, so the predictor can learn incomplete models given by computable interval restrictions on the probability of the next bit. The paper also establishes weighted forms of unbiasedness and calibration, and discusses whether the optimal predictor's conditional probabilities suffice to compute every hypothesis in the reflective class.
- The hypothesis class of ordinary Solomonoff induction has no effective enumeration, and the authors cite that it may fail to learn simple regularities on selected bits; whether the optimal predictor of the reflective oracle version can compete with effective prediction strategies that specify only part of the data-generating process is said by the paper to have been previously unanswered.
- On the reflective oracle hypothesis class, partial models are rewritten as comparison hypotheses that borrow the predictor's conditional probabilities, multiplicative domination is used to bound their cumulative log-loss advantage, and almost-sure convergence under computable interval constraints is then derived; a predictor is also constructed that specifies each hypothesis's conditional probability at designated histories.
- Purely theoretical results: the distance from the predictor to the computable interval [l(σ), u(σ)] tends to 0 almost surely under every compatible measure, with the expected sum of squared errors bounded by 2(K(l,u)+C)ln2; selected bits and weighted calibration hold as corollaries; and there exists an optimal predictor that presents the entire reflective hypothesis class.
- The authors state: whether a fixed predictor presents this class, and whether there exists a dominating member that does not present this class, remain unresolved (Section 3); the random selected-bit result does not answer the corresponding question for the ordinary Solomonoff predictor (Section 2.4); allowing model constraints to depend on the oracle is considered unnatural by the authors (Section 2.3).
Related events and news
1 events · 1 news mentions
News that mentions this paper
论文提出反射预言机版Solomonoff归纳
- AI Alignment Forumcites the paper's id or link基于反射预言机的Solomonoff归纳