A learning procedure for multisurface method of pattern separation

摘要

Mangasarian(5) has proposed an interesting method of pattern separation, which is called “multisurface method”. In the method, linear programming problems are recursively solved, and the correct classification of any disjoint pattern sets is basically possible. However the fact that linear programming problems are recursively solved leads to the result that it takes long computation times and requires much memory space of computer in use. This paper describes a learning procedure for multi-surface method instead of linear programming to avoid drawbacks above. The proposed learning algorithm requires only repetitive simple calculations. Experimental results show that computation times required by learning procedure proposed are shorter than those by linear programming.