Support Vector Machine for BlogTrader Platform
I've got the Support Vector Machine (SVM) running on BlogTrader Platform. The results are more stable than MLP and I'm still being familar with the C and sigma parameters of SVR.
SVM is a good tool that based on Kernel and Statistical Learning theory, and it can reach the globle minima, with only few parameters needed to be adjusted which may be easily applied GA.
For those who are interesting in machine learning, there are several good books are recommented:
Statistical Learning Theory [Vladimir N. Vapnik]
An intruduction to Support Vector Machines and Other Kernel-based Learning Methods [Nello Cristianini, John Shawe-Taylor]
Kernel Methods for Pattern Analysis [John Shawe-Taley, Nello Cristianini]
Posted at 10:21PM Jul 30, 2006 by dcaoyuan in AIOTrade | Comments[2]
Ok. I am just doing this to save your time. The choice of architecture really does not matter so much. Every time you minimize squared error only you will end up in sharp minimum of the error surface. Sharp minimum corresponds to complex model. Complex model correspond to low probability model. You can define randomness by complexity/incompressibility. You need to minimize some additional error term which would allow you to prune some weights for instance. Occam's razor (Principle of Parsimony). I gave you enough hints so please do some deeper research befor you continue with your AI. And still I don't see where SVM should be better than MLP. Good luck!
Posted by aa on August 01, 2006 at 06:48 AM PDT #
Hi,
Thanks for your suggestions.
I understand what your mean. As the Occam's razer, we should define what's simple or complexity. For SVM, we should choose a good C ot Nu to balance the complexity and error, and I find that in my model, when I reduce complexity (such as when the number of support vectors decreased), the generalization performance trend to be better. The complexity in SVM is defined something like structure complexity.
In MLP, I sometimes got better result than SVM, but the result is not robust, the result may be different in next training. So, I think I should try to do machine committes etc.
I'm still in finding the good parameters for my model in SVM. After that, I'll post some results to show the robust of SVM on generalization performance.
Posted by 222.248.240.163 on August 02, 2006 at 12:36 AM PDT #