This manual describes the METAL Machine Learning Experimentation Environment (METAL-MLEE for short) and how it is used in a meta-learning setting. METAL-MLEE is a set of programs, supporting files, and standards that allow the organized, self-documenting and distributed execution of machine learning experiments. The results obtained by these experiments can be used as new meta-data for the METAL Data Mining Advisor (cite: theadvisorstuff).
This manual is an abridged version of [Petrak 2002a] and contains additional information about the use of METAL-MLEE in the METAL meta-learning context.
Additional information on other components that are needed for meta-learning can be found in (cite: rankingstuff?) and (cite: DCT/GSI stuff). The main source of information for the METAL-tools in the internet is www.metal-kdd.org.
METAL-MLEE helps in obtaining the meta-data for new databases and algorithms that is needed for meta-learning. It is a set of programs and Perl-scripts that help you run the necessary experiments in an orderly fashion and create a set of standardized output files.
METAL-MLEE can be used to obtain error estimates, CPU time measurements, and other data about the performance of machine learning algorithms on a specific database. These measurements will automatically be stored in a set of output files, together with other data like database characteristics, a description of the experimentation environment and parameters, and a detailled log of the experiment.
The output files serve as the basis for meta-learning, providing the necessary data for the data mining advisor (cite: doc about local dma) or other meta-learning schemes.
A more detailled description of what the METAL-MLEE does is given in Section 3.