This is with relevance to Machine Learning and application to traffic systems and traffic flow monitoring and control in predictive model using Support vector machine with training performed by Kernel trick or Artificial Neural Network trained with back-propagation.
SVM is "local" classifier, and ANN (MLP) is "global". It means that the first use local features (kernels) at the feature space and ANN divides the feature space with hyperplanes (globally).
As well as usually we do not have the information on the separating surface, there is no preferences on use one of the classifiers. Everything depends on the expertize of modeler.
dear Taiwo
I had a research project about the eclectricity load forecast (Time serie forecast), I used SVR (several kernel) and ANN trained with back-propagation i will put the paper the link so you can see the result of the two approach.
I hope this will help
Conference Paper Parallel seasonal approach for electrical load forecasting.
SVM is "local" classifier, and ANN (MLP) is "global". It means that the first use local features (kernels) at the feature space and ANN divides the feature space with hyperplanes (globally).
As well as usually we do not have the information on the separating surface, there is no preferences on use one of the classifiers. Everything depends on the expertize of modeler.
I am disagree with Andrey's answer. According to my understanding SVM gives global optimum while ANN may give local optimum.
Krupal Parikh: You mix two different things. I agree with Andrey that when using radial basis functions we get local solution, and when we use sigmoids we get more global solution. This is different from global/local optimization procedures used in ANN and SVM. But note that the optimization problems formulated in ANN and SVM are completely different.
Taiwo: I suppose there is no answer to your question. Select the best activation function experimentally.
@Krupal , I do agree that it will depend on the type of problem. Radial base function(RBF) are more often used in conjuction with SVM while sigmoids are used in ANN. However, one may need to select the best activation function experimentally for good parameter setting of the algorithm:SVM or ANN.
Yu(2009) argues that RBF generalises well, and can be less subceptible to noise.
PS: Thank you all for the insightful comments.
Reference: Yu, H.(2011), Advantages of Radial Basis Function Networks
for Dynamic System Design, IEEE Transactions on Industrial Electronics, Vol. 58, No. 12.
.
a nice explanation of why Gaussian kernels work so fine with SVM
http://stats.stackexchange.com/questions/131138/what-makes-the-gaussian-kernel-so-magical-for-pca-and-also-in-general
.
- Badges
- Science topic