Tuesday, September 13, 2016

Speeding up Bayesian Optimizations with Random Projections - implementations -

Random projections and features used in Bayesian optimization to go faster ! Here are two implementations.

COMBO: An Efficient Bayesian Optimization Library for Materials Science by Tsuyoshi Uenoa, Trevor David Rhoneb, Zhufeng Houc, Teruyasu Mizoguchid, Koji Tsuda
In many subfields of chemistry and physics, numerous attempts have been made to accelerate scientific discovery using data-driven experimental design algorithms. Among them, Bayesian optimization has been proven to be an effective tool. A standard implementation (e.g., scikit-learn), however, can accommodate only small training data. We designed an efficient protocol for Bayesian optimization that employs Thompson sampling, random feature maps, one-rank Cholesky update and automatic hyperparameter tuning, and implemented it as an open-source python library called COMBO (COMmon Bayesian Optimization library). Promising results using COMBO to determine the atomic structure of a crystalline interface are presented. COMBO is available at https://github.com/tsudalab/combo.

Bayesian Optimization in a Billion Dimensions via Random Embeddings by Ziyu Wang, Frank Hutter, Masrour Zoghi, David Matheson, Nando de Freitas

Bayesian optimization techniques have been successfully applied to robotics, planning, sensor placement, recommendation, advertising, intelligent user interfaces and automatic algorithm configuration. Despite these successes, the approach is restricted to problems of moderate dimension, and several workshops on Bayesian optimization have identified its scaling to high-dimensions as one of the holy grails of the field. In this paper, we introduce a novel random embedding idea to attack this problem. The resulting Random EMbedding Bayesian Optimization (REMBO) algorithm is very simple, has important invariance properties, and applies to domains with both categorical and continuous variables. We present a thorough theoretical analysis of REMBO. Empirical results confirm that REMBO can effectively solve problems with billions of dimensions, provided the intrinsic dimensionality is low. They also show that REMBO achieves state-of-the-art performance in optimizing the 47 discrete parameters of a popular mixed integer linear programming solver.

The Rembo repository is here: https://github.com/ziyuw/rembo
Join the CompressiveSensing subreddit or the Google+ Community or the Facebook page and post there !
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.

No comments: