Also don't forget that you can set NumOptCycles to 2, to make 2 optimization passes. The first pass will find initial set of optimum parameters, then second pass will optimize them all once again, giving much better fit. For better, or for worse.

I must agree with the manual, setting NumOptCycles to 3 either doesn't produce any improvements (but burns even more cpu cycles) or really overfits the parameters. Avoid it.

NumOptCycles mechanism provides pretty good compromise between too simplistic one try at best parameters, and exhaustive bruteforce sweep of all possible combinations.

Oh yeah good point Acid... I did forget about that! I will have to play around with NumOptCycles more. It seems like manually finding and setting the optimal "start" value might be the best choice (since parameters are inter-dependent on each others start values) but perhaps NumOptCycles is a good compromise if you just need a quicker result.