Squint provides a codebase to perform numerical proof-of-concept experiments in learning theory, particularly in Prediction with Expert Advice problems, a core problem in learning theory.

The software has been developed in the Machine Learning group and is a companion to the research paper 'Second-order Quantile Methods for Experts and Combinatorial Games'. The algorithm is provably robust, was designed to be highly adaptive, and performs excellently in a broad spectrum of practical environments.

The MetaGrad algorithm can be regarded as a successor to Squint, in the sense that it applies to a more general problem, i.e. Online Convex Optimization.

Additional Metadata
THEME Information (theme 2)
Project Safe Statistics , Machine Learning at the Intrinsic Task Pace
Rights MIT License
Note

This repository contains a numerically stable implementation of the Squint algorithm from the paper

Second-Order Quantile Methods for Experts and Combinatorial Games Koolen, Wouter M., and Tim van Erven In Proceedings of the 28th Annual Conference on Learning Theory (COLT) 2015, 1155–75

Grant This work was funded by the The Netherlands Organisation for Scientific Research (NWO); grant id nwo/639.021.439 - Machine Learning at the Intrinsic Task Pace
Citation
Koolen-Wijkstra, W.M. (2015). Squint.