This research involves the development of a software system for automatically finding optimal or near-optimal classification methodologies for arbitrary supervised classification problems through experimentation. The system processes given taxonomies and features and outputs comparisons of the effectiveness of different classification methodologies with different parameter configurations.
A variety of classification and feature weighting algorithms are included in the system. Various techniques are evaluated in terms of both success rates and computational complexity. The system makes use of the WEKA data mining engine. The software is being designed with the goal of eventually integrating it with the D2K system to allow for distributed processing.
Although the system is designed to be able to deal with general classification problems in any domain, it is tested here using primarily musical applications. It is hoped to eventually integrate this software into the International Music Information Retrieval Systems Evaluations Laboratory (IMIRSEL) project. In particular, this software is currently being developed in parallel with research into automatically segmenting continuous audio recordings into appropriate tracks. Further tests are being conducted involving genre classification, beatbox classification and instrument identification. More general testing is also being performed using the UCI Machine Learning Repository.
Audio and MIDI feature selection sub-systems are also being developed to enable ACE to perform the full range of audio or MIDI classification tasks, if necessary.
The alpha demo version of ACE is now complete. It can be downloaded here. Note, however, that this is the initial ACE demo prototype that was constructed to study the project viability. The ACE software has progressed very significantly since then, but is currently at a stage in development where the interface is being reimplemented, so it is not suitable for release. This early protype is the last "stable" release, and it is being released only in order to give those interested in ACE a rough idea of how ACE is being implemented. This release is in no way fully stable or truly debugged. This is the release that was used to process a variety of UCI datasets in Weka ARFF format.
