if u use a learning algorithm with data, then the algorithm will modify the weights on the network in each learning epoch (e.g. stochastic gradient descent) in the direction of the steepest gradient(s). this is locally optimal, although not guaranteed to give u the global solution, but is good enough for most applications

You can train both in online as well as batch mode. In online mode you can change the weights based on the mean squared error per feature vector presented (to minimize its difference. This can be done by comparing the actual weight matrix vs the perturbed weight matrix using (GA or PSO). Due to the potential for noise what you can do is use a moving window on the Mean square error total of feature vectors presented. As you extend the window you approach the whole set and then  you are  basically doing batch learning. In batch learning you present the whole epoch and calculate the total error. After this calculation you then perturb the matrix (again using GA or PSO) run another epoch to see which provides a better result.

The questions that are open that I would try to address in this type of research are:

1) as explained above the window that you will use to perturb the weight matrix)

2) The number of entries in the weight matrix to perturb

Hope this helps

Thanku all.