camir-aes2014: toolboxes/SVM-light/Readme_optimization_relative

first hg version after svn

rev	line source
wolffd@0	1 Solving general optimization problems
wolffd@0	2 -------------------------------------
wolffd@0	3
wolffd@0	4 You can use SVM-light to solve general optimzation problems of the form:
wolffd@0	5
wolffd@0	6 min 0.5 w*w + C sum_i C_i \xi_i
wolffd@0	7 s.t. x_i * w > rhs_i - \xi_i
wolffd@0	8
wolffd@0	9 Use the option "-z o". This allows specifying a training set where the examples are the inequality constraints. For example, to specify the problem
wolffd@0	10
wolffd@0	11 min 0.5 w*w + 10 (1000 \xi_1 + 1 \xi_2 + 1 \xi_3 + 1 \xi_4)
wolffd@0	12 s.t. 1 w_1 >= 0 - \xi_1
wolffd@0	13 -2 w_1 >= 1 - \xi_2
wolffd@0	14 2 w_3 >= 2 - \xi_3
wolffd@0	15 2 w_2 + 1 w_3 >= 3 - \xi_4
wolffd@0	16
wolffd@0	17 you can use the training set
wolffd@0	18
wolffd@0	19 0 cost:10000 1:1
wolffd@0	20 1 1:-2
wolffd@0	21 2 3:2
wolffd@0	22 3 2:3 3:1
wolffd@0	23
wolffd@0	24 and run
wolffd@0	25
wolffd@0	26 svm_learn -c 10 -z o train.dat model
wolffd@0	27
wolffd@0	28 The format is just like the normal SVM-light format. Each line corresponds to one inequality. However, the first element of each line is the right-hand side of the inequality. The remainder of the line specifies the left-hand side. The parameter cost:<value> is optional and lets you specify a factor by which the value of the slack variable is weighted in the objective. The general regularization parameter (10 in the example) is specified with the option -c <value> on the command line.
wolffd@0	29
wolffd@0	30 To classify new inequalities, you can use svm_classify in the normal way.

Mercurial > hg > camir-aes2014