|Title:||Resource-Bounded Selection of Learning Algorithms
|Authors:||Assaf Glazer and Shaul Markovitch
|Abstract:||We present a new setup and a proposed solution for the resource-bounded learning-algorithm selection problem. Our setup assumes that a resource allocation pattern in a meta-learning system can be predefined. We use these predefined patterns to design novel preference utility functions between pairs of learning algorithms. These functions are exploited by a meta-learning process to support the selection of the best estimated learning algorithm when resources are limited. The resource allocation pattern is formulated as a distribution over the allocated execution time, and it may be considered more natural or more convenient in many real-life learning systems in which the pattern can be estimated or refined from previously solved learning tasks. In addition, we introduce a new evaluation methodology that uses an interruptible anytime algorithm to evaluate ranking method performance. This evaluation methodology, combined with a distribution over the allocated execution time, allows us to measure the ranking method's contribution in terms of its expected estimated error. Finally, we introduce a general anytime learning-algorithm selection framework for operating in a resource-bounded environment. Our framework can be used, for example, in a meta-learning advisory system or in an automated centralized server dealing with thousands of incoming learning tasks.|
|Copyright||The above paper is copyright by the Technion, Author(s), or others. Please contact the author(s) for more information|
Remark: Any link to this technical report should be to this page (http://www.cs.technion.ac.il/users/wwwb/cgi-bin/tr-info.cgi/2013/CS/CS-2013-06), rather than to the URL of the PDF files directly. The latter URLs may change without notice.
To the list of the CS technical reports of 2013
To the main CS technical reports page
Computer science department, Technion