The randomness (and exploration) encouraged by batch training also helps avoid 'real' minima, if they exist.