PHD2 implements a hysteresis algorithm where you can specify how much the next correction sent relies on the most recent sample versus the history of samples.

There are also several options which use a low pass filter to achieve a similar end.

Also you can specify a maximum move per step.  You can also tell if if is see a very large excursion (due to a dither move) it will attempt to recover more quickly than the max move you specified.

Most is summarized here:

http://openphdguiding.org/man/Guide_algorithms.htm

I've found using long (4-6 second) exposures extremely useful when dealing with nights of poor seeing.   It is a simple experiment to incrementally increase the exposure time and watch the random scatter of samples of the DEC axis while guiding is turned off to gauge how well this works. 

If seeing conditions are so bad that you are seeing excursions more than say 1 arcsecond regularly it is probably time to shut down as that night is a lost cause.

Something to consider is there are other factors that can cause large jumps which have nothing to do with mechanical or atmospheric behavior.  For example, for a period apparently some QHY cameras with certain QHY drivers would occasionally output frames that were shifted by 1 pixel!  Imagine the consequences on guiding when using a smaller guide scope and a pixels scale of several arcseconds/pixel.

Michael Fulbright