So you have a dependency of 3 values (deviation in x, y, and z) from 4 parameters (position in x, y, and z as well as load F). You might also want to consider the movement vectors (direction and speed) for input, since I suspect that the deviation(x, y, z) will be a swinging motion, not static!
In any case, while approximation algorithms for functions in one (curves) or two variables (surfaces) are fairly well established, I am not aware of a generic algorithm for functions of more than two variables. Of course, mathematically the problem is not that different, and it's possible to generalize from approximation algorithm for functions of two variables to functions of 3 or more, but it may be difficult to find readymade program code for that.
More importantly, if you are not familiar with the required mathematic background, you will have a hard time of setting up a system and fitting it to your requirements.
For these reasons, your best approach may actually be something entirely different: train an Artifical Neural Network (ANN) to the task! An ANN has an arbitrary number of input units (one for each input value) and an equally arbitrary number of output values (one for each required/expected output). There are several articles on ANNs on this site: you may want to search for the keyword "backpropagation" - I think backpropagation networks may be best suited to your problem.
You could set up a network with 3 layers of neurons: the first should contain one unit for each input: x, y, z, and posssibly the 3 components of your movement vector (lenght corresponding to speed). The third layer contains one unit for each output: dx, dy, dz (and maybe 3 more for movement: vx, vy, vz). The second layer can have any number of units, but you need at least as many units as you have either inputs or outputs (i. e. the min of the two numbers) to prevent loss of information (in fact I would advise to use the max of the two numbers or more - IME that provides faster training and better results).
Unfortunately I haven't been using ANNs since the 90s, so I can't offer more advice at this point. But the articles on backpropagation networks on this site may prove to be helpful.
P.S. - my reasons for suggesting an ANN:
As mentioned above, you could probably find an optimized algorithm for this kind of problem, if you search for it long and hard enough. However, I am often confronted with similarly hard problems and my experience is that (a) what you find never really fits your problem (so you have to work to make it fit), (b) obtaining the actual code either requires payment (usually not a big problem unless we're talking of full blown math libraries), licensing (more often a problem than not, especially because many 'open' licenses are incompatible with use in industrial environments), or you get only the binaries rather than the source code (which makes it a lot harder to adapt according to (a)), or (c) isn't described sufficently to tell if it even solves your problem. More often than not it's all three.
The benefits of ANNs are:
1. Plenty of free implementations around
2. Requires little knowledge of math
3. can be easily adapted to changing requirements or new inputs and insights
All these appear to be relvant in your case.
The disadvantages are
1. Takes a lot of input data and time to train properly
2. You have very little influence over the shape of the solution space
The first may not be very important to you, as you can take an arbitrary number of input samples, and you only have to train once. It would be a problem if generating the input takes a lot of time or is costly.
The second is not a problem if you don't even know about how to influence the solution shape in other algorithms (e. g. spline approximation), or what would be a good shape in the first place. In my experience ANNs provide rather smooth shapes, and I believe that's good for you.
Updated 5-Jun-12 22:56pm
v3
A shot into the dark:
1) build the physical machinery
2) make sure you can move to all needed positions
3) make sure you can measure accurately enough the controlled position
4) make sure you can measure accurately enough the actual position
5) calibrate that system with a bunch of loads from a feather to the max load
5.1) take a load
5.2) move to all possible positions within a grid of suitable granularity
5.3) measure the deviation from the ideal position
5.4) store that x,y,z,F --> delty_x, delta_y, delta_z tables
6) once you have accurate enough data, interpolate between the measured points
while on duty
If this is any good solution for your problem is completely unclear to me. You have to decide if such an approach is feasible at all.
Please state your problem more clearly as Sergey, Manfred, and Stefan already pointed out.
Cheers
Andi
PS: It does not calculate any analytical curve or shape, though.
Updated 4-Jun-12 8:45am
v3