Do the GA- or PSO or ABC- based BPANN (LM or BR) perform better than the traditional BRANN in terms of function approximation? My answer is ' Not Necessarily'. How about your opinion? Any discussion would be welcome. Thax.
That is a very mysterious question. I can't understand the question, so I have no hope of giving an answer.
Sorry about that. I will clarify this question again. As is known to all, the traditional Back Propagation Artificial Neural Network(BPANN) is prone to reach a local minimum. That is to say BPANN is prone to cause the overfiting problem. Therefore, several integration methods, e.g. Genetic Algorithm based BPANN, have been developed to avoid the overfiting problem and improve the accuracy of the traditional BPANN. Furthermore, several international journal papers demonstrated these integration methods perform better than the traditional BPANN. However, for my specific problem, i.e. flammable cloud size estimation, the conclusion is questionable. For my problem, I find that Beyesian Reguralization Aritificial Neural Network ( BRANN) exhibits better generalization capacity compared to those integration methods especially under limited trained data and it is easier for BRANN to determine the optimal hidden neurons . However, I find that those integration methods are more suitable for those problems with large number of data( more complicated problem) since those methods can significantly improve the accuracy of the traditional BPANN or BRANN.
The question is clearer. Sorry, i cannot answer it. I am a geologist.