Core Concepts
It is possible to achieve near-optimal online estimation error by converting offline estimation algorithms into online estimation algorithms in a black-box fashion, up to logarithmic factors.
Abstract
The paper introduces a new framework called Oracle-Efficient Online Estimation (OEOE), where the learner can only interact with the data stream indirectly through a sequence of offline estimators produced by a black-box algorithm.
The key results are:
Statistical complexity:
For finite classes F, there exists an oracle-efficient algorithm that achieves near-optimal online estimation error, up to a logarithmic factor (Theorem 3.1). This is shown to be nearly tight (Theorem 3.2).
For memoryless oracle-efficient algorithms, strong impossibility results are shown (Theorem 3.3).
A general reduction from oracle-efficient online estimation to delayed online learning is provided (Theorem 3.4), and used to characterize oracle-efficient learnability for classification with infinite classes (Theorem 3.5).
Computational complexity:
Under standard complexity assumptions, there do not exist polynomial-time algorithms with non-trivial online estimation error in the OEOE framework (Theorem 4.1).
However, for conditional density estimation, online estimation in the OEOE framework is no harder computationally than online estimation with arbitrary, unrestricted algorithms (Theorem 4.2).
The results have implications for interactive decision making, showing that it is information-theoretically possible to achieve near-optimal regret for any interactive decision making problem using an algorithm that accesses the data stream only through offline estimation oracles (Corollary 5.1).