Pearce, Michael and Branke, Jürgen (2017) Efficient information collection on portfolios. Working Paper. Coventry: University of Warwick.

Preview

PDF (Working paper) WRAP_0871351-wbs-260117-jorv4.pdf - Other - Requires a PDF viewer. Download (5Mb) | Preview

Request Changes to record.

Abstract

This paper tackles the problem of efficiently collecting data to learn a classifier, or mapping, from each task to the best performing tool, where tasks are described by continuous features and there is a portfolio of tools to choose from. A typical example is selecting an optimization algorithm from a portfolio of algorithms, based on some features of the problem instance to be solved. Information is collected by testing a tool on a task and observing its (possibly stochastic) performance. The goal is to minimize the opportunity cost of the constructed mapping, where opportunity cost is the difference between the performance of the true best tool for each task, and the performance of the tool chosen by the constructed mapping, summed over all tasks. We propose several fully sequential information collection
policies based on Bayesian statistics and Gaussian Process models. In each step, they myopically sample the (task, tool) pair that promises the highest value of the information collected. We prove optimality under certain conditions and empirically demonstrate that our methods significantly outperform standard approaches on a set of synthetic benchmark problems.

Item Type:	Working or Discussion Paper (Working Paper)
Divisions:	Faculty of Science, Engineering and Medicine > Research Centres > Centre for Complexity Science Faculty of Science, Engineering and Medicine > Science > Mathematics Faculty of Social Sciences > Warwick Business School
Publisher:	University of Warwick
Place of Publication:	Coventry
Official Date:	26 January 2017
Dates:	Date Event 26 January 2017 Available
Institution:	University of Warwick
Status:	Peer Reviewed
Publication Status:	Published

Request changes or add full text files to a record

Repository staff actions (login required)

View Item

Downloads