looms: leave-one-out model selectionlooms uses a  slightly modifiedBSVM 1.1 to perform model selection on binary classification problems. Currently the RBF kernel is supported.*****************************************************************           COPYRIGHT NOTIFICATIONlooms can be freely used for research purpose.Use for commercial purposes is expressly prohibitedwithout contacting the authors.looms is provided "as is" without express or implied warranty.Jen-Hao Lee and Chih-Jen LinDepartment of Computer Science andInformation EngineeringNational Taiwan University Taipei 106, Taiwane-mail: cjlin@csie.ntu.edu.twhttp://www.csie.ntu.edu.tw/~cjlin/loomslooms is based on the SVM software BSVMby Chih-Wei Hsu and Chih-Jen Lin.Please read the COPYRIGHT NOTIFICATIONbefore using it.The BSVM homepage is in http://www.csie.ntu.edu.tw/~cjlin/bsvm*****************************************************************           INSTRUCTIONS          1. Create the looms directory structure with        unzip looms.zip   This produces the directory looms and several subdirectories.2. Change directories to looms and install looms with         make 3. Use -h to know options of looms		looms -h 	4. Run the sample problems by executing looms. A model    with the best loo rate is generated	looms heart_scale heart_model > heart_out5. (optional) Use the following perl script to extract   loo rates of all models in a matrix form:	extr_rate.pl heart_out   Users can then use software such as MATLAB to draw 3-Dsurfaces (e.g. the picture on the looms homepage)6. Test the classifier	classify heart_scale heart_model classfied_result    test data: heart_scale   support vectors: heart_model   classified_result: results after classification *****************************************************************           FILE FORMATS and PARAMETERSThe file format is the same as that of BSVM. Please referto README.BSVMType looms to know the usage of looms:Usage: looms [options] training_set_file [model_file]options: -a aclb : use early stop method with accuracy lower bond aclb -c cost : set cost C of constraints violation (default 1)    cost := { c | c1-c2{x|+|-}c_prog }  ( no space allowed )    ( default 1-1024x2 ) -et epsilon : full training termination criterion tolerance (default 0.001) -el epsilon : loo termination criterion tolerance (default = 0.1) -g gamma : set gamma in kernel function (default 1/k)    gamma := { g | g1-g2{x|+|-}g_prog }  ( no space allowed )    ( default 0.00025-2.048x2 ) -h : this help -m cachesize : set cache memory size in MB (default 160) -q qpsize : set subproblem size (default 10) -s on/off: turns on/off S-shape sequence (default on) -v {0,1,2,3} : verbosity (default 1)    0 -- result of the whole model selection    1 -- brief information on the loo of each parameter pair    2 -- information on each loo phase    3 -- information on each iteration of optimization At the present time, the search space must be a rectangle on (C,gamma) space.Examples:		% looms -c 1-64x4 -g 0.02-1.28x4 -a 0.7 -s -et 0.001 -el 0.1 heart_scaleWe try models with C = 1, 4, 16, 64 and gamma = 0.02,0.08, 0.32, 1.28 and see if there are any models with loo accuracy >= 0.7. Therefore, the loo calculation ofeach model stops earlier if its accuracy is already < 0.7. A loose stopping tolerance -el leads to faster calculationof loo (see Section 3 of the looms paper). Note that -s and -a are speeding methods in this implementation where -s is activated as the default. For details of these two options, please refer to Section 4 and 5 of the looms paper, respectively.*****************************************************************           ADDITIONAL INFORMATIONJ.-H. Lee and C.-J. Lin Automatic model selection for support vector machineshttp://www.csie.ntu.edu.tw/~cjlin/papers/modelselect.ps.gzAcknowledgments:This work was supported in part bythe National Science Council of Taiwan via the grantNSC 89-2213-E-002-013.The authors thank Chih-Wei Hsufor many helpful discussions and comments.