Meta-learning from experiment databases: An illustration

Gaining insights into the behavior of learning algorithms generally involves studying the performance of the algorithms on many different datasets, with various parameter settings, and analyzing the results. A useful method to learn effectively from previous learning episodes are experiment databases: databases designed to store detailed descriptions of a large number of experiments, selected to cover a wide range of conditions. We illustrate how such databases allow us to easily gain insights into a wide range of questions on algorithm behavior by just querying them and interpreting the results, or by using data mining methods to automatically discover patterns in learning algorithm performance. By putting these databases online, they serve as a repository of experimental results that can be (re)used by various researchers to easily obtain new insights.status: publishe