Issued - Nov 1, 2012
Updated - Nov 2, 2016
This package contains the algorithm for learning the Augmented-SVM classifier function for combining multiple non-linear dynamics. The algorithm was presented in the paper:
Shukla, A. and Billard, A. "Augmented-SVM: Automatic space partitioning for combining multiple non-linear dynamics." Neural Information Processing Systems (NIPS) 2012. Tahoe, Nevada.
Useful for: Automatic Space Partitoning of multiple non-linear dynamics
This package is organised as follows.
- The root folder contains cpp source files in
src/andinclude/which are compiled using cmake into a librarylib/ASVMLearning.soThis is then linked to the executablebin/trainwhich is a command-line interface to the learning algorithm. - The
matlab/folder contains mex interface to the main library and the functionsvmtrainfromlibsvm. - The ASVM utility functions are in
matlab/SVMUtil. - A GUI is provided in
matlab/SVMWidgetto draw trajectories using a mouse or stylus and call the various learning algorithms to get the A-SVM model. This is the calling point for all thecpp/mexfunctions compiled before.
Extract the A-SVM folder to any location.
>> cd <A-SVM_root_dir>
>> mkdir build
>> cd build
>> ccmake ..
>> make
You may optionally choose NLOPT at the ccmake gui. You must have nlopt "make install"-ed somewhere on you filesystem. If you chose standard install location for NLOPT, ccmake will find the required files automatically. If you installed NLOPT in some other location you may need to manually give the location of the nlopt library and nlopt.hpp in the ccmake gui. Also, in any case, do not forget to add the folder containing the nlopt library to LD_LIBRARY_PATH.
You can also choose to switch off the NLOPT option and in that case you will not be able to use the NLOPT algorithms from the matlab gui.
Once ASVMLearning.so is built, you need to compile the mex files from matlab. For that just run the following command in your matlab command line
>> setup_path.m
>> make
This will take care of compiling all the mex interfaces and adding relevant folders to matlab path. If you get GLIBC_XXX errors on calling the mex functions, you may try to run matlab using the provided script "run_matlab.sh".
To run the matlab widget call:
>> svm_widget.m
Extract the ASVMLearning folder to any location.
>> Create a build directory inside ASVMLearning
>> run cmake-gui with source_folder=ASVMLearning and build_folder=ASVMLearning/build
>> configure
>> Optionally choose NLOPT (See above LINUX installation for details).
>> generate
You must compile in the "Release" version otherwise the mex linkage will fail later on.
Once ASVMLearning.so is built, you need to compile the mex files from matlab. For that just run matlab from the ASVMLearning root folder and run "setup_path.m". This will take care of compiling all the mex interfaces and adding relevant folders to matlab path.
This package uses the function svmtrain from Libsvm. It contains a modified subset of the original libsvm source which additionally returns the indices of the chosen support vectors within the model. Full version of Libsvm is available at http://www.csie.ntu.edu.tw/~cjlin/libsvm Please read the COPYRIGHT file before using Libsvm.
This package gives access to using all the NLOPT algorithms for learning the A-SVM model. NLOPT is available at http://ab-initio.mit.edu/wiki/index.php/NLopt
This package can optionally use the IPOPT solver if it is compiled with its MATLAB interface. IPOPT is available at https://projects.coin-or.org/Ipopt
Current Maintainer: Nadia Figueroa