Some enhancements of the constraint based decomposition training architecture

This paper presents three mechanisms: locking detection (LD), redundancy elimination (RE) and generalisation control (GC). Although originated, implemented and tested for the constraint based decomposition (CBD), LD and RE address general drawbacks of constructive algorithms and can be applied to other such algorithms, as well. LD stops the training of a hyperplane if its position is pin-pointed by nearby patterns, thus improving the training speed. RE reduces the number of hyperplanes used in the solution. RE works on-line during the training, thus eliminating the need for a separate pruning stage. Finally, GC addresses the generalisation properties of the solution. It is shown that CBD's generalisation can be controlled by the user through the ordering of the pattern set. The experiments presented show that these mechanisms are effective on various types of problems.