I do not know exactly  what your question is about. 

There is not an exact method to find the hyper-paramteres (C, epsilon and possible kernel parameters)  and most of the time a simple Grid Search is applied. However, this search is very time consuming, so it is better to use an optimzation search algorithm (EA, PSO, etc) which reduces the training time and (with a good search algorithm implementation ) improves the accuracy. The optimization algorithm will create several models with different hyper-parameters configuration and a training set. Each model is evaluated with the training set or a validation set (in case you use a validation method, highly recommended) which gives an error for that model. The algorithm will iterate generating more models until this converges to a good combination of hyper-parameters. This does not mean that the algorithm will find the best hyper-parameters, it does not test all the possible combination, but it will give you one good enough.

I hope this answer will help you.

PD: you should use always a validation method if you want to avoid overfitting problems.

Dear Lokesh Gagnani ,

The performance of SVM mode depends on the parameters. PSO is a heuristic approach, and used for optimization or searching problem.  So, in your work, you consider PSO for finding the optimal value of C and epsilon.

You follow this paper "Hybrid functional link artificial neural network

approach for predicting maintainability of

object-oriented software". In this paper, we described the step of PSO.

My Question is how to OPTIMISE the parameters of SVM using PSO.

Hi,

is your question of a technical nature (how to use Java to do what you want to?

There is a framework called JSwarm that lets you implement PSO very easily. The more tricky part is integrating SVM-Java-Implementation. There are few implementation (libSVM has a Java-interface in Weka as far as I know, also Rapifminer should provide something)

Or is your question of a theoretical nature (how to do it)?

You have to implement the objective function f(x1,...,xn) of the PSO (that is going to be minimized). This objective function has to conduct a training of the SVM based on the parameters x1...xn which are interpreted as SVM parameters (bandwidth and so on). The SVM that has been trained with theese parameters is evaluated and the result of this evaluations (the quality of the SVM) is the result of the objective function. In this way PSO will vary the parameters until a good set has been found. The absolute optimum will most likely never been found.

Best regards.

Dear Lokesh Gagnani, 

You may use PSO, GA, CSA etc algorithm for parameter optimization of SVM.  In JAVA, it is little bit difficult to implement PSO for SVM parameter optimization. You may use MATLAB tool. Please follow these paper. In these paper, we use PSO, GA, and CSA for finding optimal weight of neural network. You may also use same steps for SVM parameter optimization.

https://www.researchgate.net/publication/299328355_Hybrid_functional_link_artificial_neural_network_approach_for_predicting_maintainability_of_object-oriented_software

https://www.researchgate.net/publication/282267465_Neuro_-_Genetic_Approach_for_Predicting_Maintainability_Using_Chidamber_and_Kemerer_Software_Metrics_Suite

https://www.researchgate.net/publication/282267465_Neuro_-_Genetic_Approach_for_Predicting_Maintainability_Using_Chidamber_and_Kemerer_Software_Metrics_Suite?ev=prf_pub

Article Hybrid functional link artificial neural network approach fo...

Article Neuro – Genetic Approach for Predicting Maintainability Usin...

Dear Lokesh,

What platform do you want to work in? JAVA or MATLAB?

If you want to work with MATLAB, you should download psomatlab toolbox from following URL.

https://code.google.com/archive/p/psomatlab/downloads

Its usage is exactly like ga() function in MATLAB. If you have any problem in using this toolbox, please contact me.

For optimizing ANN parameters (weights and biases) using PSO, codes in MATLAB is available in the following article of mine. Actually, I do not have any idea about SVM to convert the codes for the same.

https://www.researchgate.net/deref/http%3A%2F%2Fdx.doi.org%2F10.13140%2FRG.2.1.2579.3524

My objective function takes SVM parameters and we gonna trry to minimoze theclassification error as follow:

f

You need to implement an algorithm that alters the values of C and Gamma for your objective function. If you are going to use PSO, following URL gives you an easy to use MATLAB code.

http://yarpiz.com/50/ypea102-particle-swarm-optimization

It is notable that if your dataset is huge, this search will be time consuming (SVM calculation is the bottleneck); because you need N populations for which different values for C and Gamma are set, so SVM fitting is done N times for each iteration. Also you may need M iterations to minimize your objective function. However, multi-threading / parallel computing decreases the time required drastically.

I have implemented this method with a GA-based feature selection method; so you can imagine how long it takes to find proper variables and optimize C and Gamma !!

i need PSO-SVM vode in matlab.

https://github.com/sli1989/svm-parameter-selection