DocumentCode :
2962131
Title :
A dynamic system framework for the decomposition method solving support vector machines
Author :
Lai, D. ; Mani, N. ; Palaniswami, M.
Author_Institution :
Dept. of Electr. & Comput. Syst. Eng., Monash Univ., Clayton, Vic., Australia
fYear :
2004
fDate :
14-17 Dec. 2004
Firstpage :
283
Lastpage :
288
Abstract :
The decomposition method is generally used to solve the quadratic program of support vector machines. The rate of convergence of this method is largely dependant on the sequence of sub-problems solved. In order to study ways of increasing the convergence, we propose a dynamic system perspective to model the dynamics of the decomposition method. In particular, the minimization of a sub-problem can be viewed as an autonomous dissipative system in terms of second order differential equations. The gradients of the sub-problems and the inequality constraints are explicitly modelled as system variables. Using these models, we then define a general decomposition method as a non-autonomous system composed of sub-systems that operate for discrete time intervals. The dependance of this system on time is depicted by a time dependent permutation matrix which functions as an indicator for operating subsystem components.
Keywords :
convergence of numerical methods; differential equations; learning (artificial intelligence); matrix algebra; minimisation; pattern recognition; quadratic programming; support vector machines; autonomous dissipative system; convergence rate; decomposition method; discrete time intervals; dynamic system framework; inequality constraints; minimization; nonautonomous system; quadratic program; second order differential equations; sub-problem gradients; support vector machines; system variables; time dependent permutation matrix; Convergence; Differential equations; Lagrangian functions; Linear matrix inequalities; Matrix decomposition; Pattern recognition; Supervised learning; Support vector machine classification; Support vector machines; Systems engineering and theory;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Intelligent Sensors, Sensor Networks and Information Processing Conference, 2004. Proceedings of the 2004
Print_ISBN :
0-7803-8894-1
Type :
conf
DOI :
10.1109/ISSNIP.2004.1417476
Filename :
1417476
Link To Document :
بازگشت