training NARX multiple times randomly changing starting parameters from the previous time

I'm using NARX for identifying the model in MIMO input-output time series. Following the suggestions, I train first the open loop net, then I transform it in closed loop and retrain. Usually performance in closed loop is much worse than in open loop, then I implemented the following expedient: I repeat training several time, the best result is always saved, and each time the initial parameters are randomly moved from the previous best result. Time to time the span of the initial parameter perturbations is reduced. In this way I obtain great improvements, measured from the relative standard deviation of the fitting output error. But the process is slow. Does a better way of doing exist? Giuseppe