Современные информационные системы / 2. Вычислительная техника и программирование

 

Kryuchin O.V., Rakitin A.A.

Tambov State University named after G.R. Derzhavin, Russia

The information processes for the function coefficient minimization efficiency increasing using quasi-Newton methods

 

In optimization theory quasi-Newton methods are algorithms for finding a local maximum and minimum of function

(1)

where  is the minimized vector. Quasi-Newton methods are based on the Newton method to find the stationary point of a function (1), where the gradient is equal to 0. Newton method assumes that function (1) can be locally approximated as a quadratic in the region around the optimum and uses the first

(2)

and the second derivatives

(3)

to find the stationary point. In formulas (2)-(3)  is the vector  size. In quasi-Newton methods Hessian matrix (3) to be minimized does not need to be computed. The Hessian is updated by analyzing successive gradient vectors instead. Quasi-Newton methods are a generalization of the secant method to find the first derivative root for multidimensional problems. In multi-dimensions the secant equation is under-determined, and quasi-Newton methods differ in how they constrain the solution, typically by adding a simple low-rank update to the current estimate of the Hessian [1].

This paper aim is to develop quasi-Newton method which can be used with the large information resource elements (IR-elements) number and to do experiments for checking the developed algorithms efficiency.

As in Newton method

(4)

thus we can calculate new vector  values by formula

(5)

where  is the approximate Hessian inverse [2].

If we have n IR-elements (for example a computer cluster or a network computing) so as it done by the classic parallel gradient algorithms one IR-element (which will calling “lead”) calculates  elements of a gradient and a gradient changing and other IR-elements calculate  elements [3]. We can divide matrix  similarly and so we can develop few information process steps:

1.     calculating parts of gradient , gradient changing  and matrix  by the different IR-elements and passing the calculated gradient and the gradient changing by non-lead IR-elements to the lead;

2.     passing gradient changing by the lead IR-element to other;

3.     each non-lead IR-element calculates matrix

(6)

rows and vector

(7)

elements; the lead IR-element calculates scalar

(8)

4.     passing calculated matrix  and vector  elements by non-lead IR-element to the lead;

5.     passing formed matrix  and vector  elements by the lead IR-element to non-lead;

6.     the lead IR-element calculates

,

(9)

(10)

and

;

(11)

at that time other IR-elements calculate

,

(12)

(13)

and

;

(14)

7.     non-lead IR-elements sends to the lead matrix ;

8.     the lead IR-element calculates matrix  and new weights and sends it to other IR-elements.

For checking the developed algorithms efficient we have done an experiment. The efficiency coefficient which was calculated by formula

(15)

(where t is serial algorithm time expenses and  is parallel algorithm time expenses with n IR-elements usage) is about 80% So we can see that parallel quasi-Newton algorithms are very efficient and can be used for the function minimization.

 

Reference:

1.     William C. Davidon, Variable Metric Method for Minimization // SIOPT, 1991 – Vol. 1, Issue 1 – P. 1-17.

2.     Fletcher R. A New Approach to Variable Metric Algorithms // Computer Journal, 1970 – Vol. 13, Issue 3 – P. 317-322.

3.     Kryuchin O.V., Arzamastsev A.A., Troitzsch K.G. A universal simulator based on artificial neural networks for computer clusters // Arbeitsberichte aus dem Fachbereich Informatik , № 2/2011, http://www.uni-koblenz.de/~fb4reports/2011/2011_02_Arbeitsberichte.pdf