I have split my dataset into training and testing. My model is fix which is give same result in every running and give a good accuracy. I want to use my neural network model coupled with genetic algorithm to select the best features from 400 features. I have done two ways first way I used my root mean square error for training as fitness function. then, after the result is identical, i tested the model another way is that i trained and tested my model so that the fitness function is the testing accuracy I find better result in second way which i got higher accuracy
my question is it logical to do the second way? what does it mean if I have two values. First one is higher accuracy than second one in training but in prediction second one is higher than first one ? which one I choose ?