Using the Bees Algorithm to optimise a Support Vector Machine for wood defect classification
This paper describes a new application of the Bees Algorithm to the optimisation of a Support Vector Machine (SVM) for the problem of classifying defects in veneer wood. The algorithm, which is a swarm-based algorithm inspired by the food foraging behaviour of honey bees, was also employed to select the components making up the feature vectors to be presented to the SVM. The objective of the work was to find the best combination of SVM parameters and data features to maximise defect classification accuracy. The paper presents the results obtained to demonstrate the strengths of the Bees Algorithm as an optimisation tool.
| Attachment | Size |
|---|---|
| BeeSVM.wmv | 5.8 MB |
Hi,
How long did it take to carry out your double optimisation procedure? (Remember to specify the machine and software that you used.)
Thanks.
M R O Dent
Thanks for your question Dr. Dent. Actually it tooks 3495 sec = 58.25 minutes to complete a double optimisation. Software used: Library Support Vector Machine (LIBSVM) version 2.83 and the Bees Algorithm.
Afshin , thanks for your question,
Actually one of the reason is that Artificial Neural Networks (ANN) use Empirical Risk Minimisation (ERM), while SVMs use a Structural Risk Minimisation (SRM). For more information the comparison between ANN and SVM, visit: http://www.svms.org/anns.html
Hi,
Well done Zaidi, thats very nice results with this quite difficult data set.
Well, I don't understand the gamma variable representing the bees. Could you explain a bit more please?
Best regards,
Ebubekir
Hi Ebubekir,
I am responding in the place of Zaidi who, I believe, is away collecting nectar in sunny Malaysia. You can imagine poor Zaidi sweating hard there.
Gamma is one of the two variables we are trying to find the optimum value for. The other variable is C.
Together, a given pair (gamma, C) represents a particular bee.
I trust this is now clear to you and other readers of our paper.
Best wishes.
DTP.
Weldone Zaidi, nice paper and presentation. As far as I can recall, 93% accuracy is the best achieved acurracy in wood fault detection.
I just wonder why SVM generally gives better classification results than other methods. I mean is it because of it's structure or there are other reasons.
Many thanks,
Afshin