Operator learning for a problem class in a distributed peer-to-peer environment

Abstract

This paper discusses a promising new research direction, the automatic learning of algorithm components for problem classes. We focus on the methodology of this research direction. As an illustration, a mutation operator for a special class of subset sum problem instances is learned. The most important methodological issue is the emphasis on the generalisability of the results. Not only a methodology but also a tool is proposed. This tool is called DRM (distributed resource machine), developed as part of the DREAM project, and is capable of running distributed experiments on the Internet making a huge amount of resources available to the researcher in a robust manner. It is argued that the DRM is ideally suited for algorithm learning.