This is a follow-on posting. I've produced the Processing software to check out the two averaging algorithms and the program is up there on the web at http://www.openprocessing.org/sketch/17533413
It should be interactive, I got a bit carried away with it, allowing different sample size and noise levels using the f/F and n/N keys but on the Openprocessing site this doesn't work, something to do with it being a Javascript version, to get the fully interactive version simply copy and paste the code into your Processing IDE. It's worth doing as it's revealing to see the effects of changing sample size for different noise levels.
Anyway the bottom line is that the 'long' algorithm (red) responds much quicker to changes in load cell than the 'short' algorithm (green). However both do much the same job of filtering out noise, and there is no error build up using the 'short' version. If anything the 'short' version filters a bit better, after all it is slower to respond to variations in input signal. For the load cell, as changes are very slow and response time is not an issue, I think I'll be sticking with the short version.
One thing I can't figure is why, on the curve shown which is for filtering over 20 samples, it takes 40 samples for the red curve to hit its average. If anyone can see an error in the code I'd appreciate knowing where it is! Anyway it doesn't detract from the general conclusion.
Geo