Ideally, parameter optimization proceeds by a set of steps in parameter space. The size and direction of these steps would be based on the gradients and local curvature of the parameter Hessian. Unfortunately however, the surface is so non-parabolic that a restraining force needs to be added. This limits the size of the steps. This restraining force adds a parabolic potential to the parameter space, with the parabola centered at the current values of the parameters. As the calculation proceeds, the size of this potential rises or falls, depending on the direction the parameters move in. If, on a given cycle, they move in more-or-less the same direction as in the previous cycle, then the restraining potential is too high, and it is decreased. Otherwise it is increased.
If PARAB=n.n is present, the default restraining potential, 10000, is replaced by n.n, and not changed during the optimization.