The model you have to chose is dependent from the experimental variogram. The residuals between experimental and theoretical variogram should be minimized.

I think it would be of advantage if you post the experimental variogram.

Gaussian is the best because it follows  Normal Probability Distribution fuction...

@ Mr. Ahmad: Doesn't the Gaussian model require a normal distribution of your measurements? Lots of measured values show a lognormal distribution (e.g. geochemical and geotechnical values). So I doubt that you can generally say that the Gaussian model is the best. The advantage of the Gaussian model is that it produces smooth surfaces. So it can be the best but it doesn't have to.

I agree with Firoz Ahamad. The main problem, however, is that arcgis is not the best software for this kind of analysis. It provides very basic tools.

Thanks everybody for your valuable inputs. Meanwhile,,,@ Sergio Noce: Any better alternative please??? Waiting for the best

If I were you, I should prefer Gaussian, 

Thanks for your valuable inputs. But I need to know the reason or logic behind the model selection. In other words, what model properties make it a best fit model?? I mean, if it were Gaussian, then why?? Why rest others not?? There must be something pertaining to a particular model that makes it best fit over others at one point of time, while it may not be the same due to some other reason in a different scenario. Can somebody throw more light on this 'reason'?

 i fully endorse the statement of  Ali Yildizel 

As colleague Günther pointed out, there is no generally "good" model. It always depends on your setting and the nature and distribution of your data. For details on how to select the right model please see:

http://desktop.arcgis.com/en/analytics/case-studies/an-appropriate-kriging-model-for-soil-contamination.htm?lg=en

Thanks, Mr. Thomas

Your question is not well posed, what are your criteria for best fit? If you are trying to ensure that the interpolated values match the "true" values that is impossible since you don't know the true values (if you do there would be no need to do interpolation"

It is incorrect to refer to "kriging model" unless you are referring to the different statistical assumptions, e.g. Simple kriging refers to a specific form of the interpolator and also to specific statistical assumptions;second order stationarity and the use of a covariance function (as opposed to a variogram) and a known constant mean. In contrast Ordinary krigng is based on intrinsic stationarity (which is weaker than second order), the use of either a covariance function or a variogram and an UNKNOWN constant mean. Universal kriging is based on intrinsic stationarity, the use of a variogram and an unknown non-constant mean that is a polynomial function of the position coordinates. The equations for the coefficients in the respective kriging estimators are obtained as the solution of a system of linear equations which is the result of two constraints (the estimator is unbiased and the kriging estimation variance is minimized.

Irrespective of which form of kriging you use you must use the data to estimate/model the variogram or covariance function. Before you attempt this you need to do various exploratory analyses such as

1. Compute and plot histograms of the data values, while there are no distributional assumptions (normality is not a requirement) it is somewhat desirable for the histogram to be approximately symmetric. Non-linearly transforming the data creates other problems so in general don't do that. For some variables, e.g. hydraulic conductivity there is some justification for a logarithmic transformation. Examine the data set for "unusual" data values, ask whether there may be an explanation for them.

2. Make plot of data locations and code each with the associated data value, does there appear to be any directional effect and/or trend. Fit a trend surface to the data

3. Compute and plot both directional and omni-directional sample variograms e.g. for 0 degrees, 45, 90, 135 with 22.5 degree tolerances. If your data set is small the directional sample variograms may be very hard to interpret. The total number of data location pairs is completely determined by the total number of data locations. For both directional and omni-directional sample variograms you need to choose lag spacings and the maximal distance (about half of the maximal inter data location spacing). Ideally you want a small lag spacing but then the number of pairs per plotted point will be small, you will likely have to experiment with these choices.

4. You are using the sample variogram plots to determine the shape of the variogram model and to estimate the parameters (some software will allow you to use one or more forms of least square fitting to determine the coefficients but only after you have chosen a particular variogram model, e.g. power, spherical, exponental, gaussian, etc). A variogram must satisfy certain conditions which are not easy to test so you only use the models incorporated in the software. It is wrong to think of the different variogram models as "sub-models",they are not. Don't assume that you can choose the "right", "perfect", or "optimal" variogram model.

5. If the data suggests a non-constant mean you have to choose between Ordinary kriging and Universal kriging and that creates some additional difficulties

6. Now use "cross-validation" and you can generate about six different statistics to help decide between variogram models and also coefficients in the model

See  1991, Myers,D.E., On Variogram Estimation in Proceedings of the First Inter. Conf. Stat. Comp., Cesme, Turkey, 30 Mar.-2 April 1987, Vol II, American Sciences Press, 261-281 for a discussion of the cross-validation statistics)

1991, Myers,D.E.Interpolation and Estimation with Spatially Located Data Chemometrics and Intelligent Laboratory Systems 11, 209-228  for a tutorial on variogram estimation/modeling.

The cross validation results may send you back to step 1

Interpolation with Radial Basis functions is theoretically equivalent to Universal kriging.

If the data has already been collected (the most common situation), the data set may not be sufficient to get good results, e.g. you don't want the data locations on a regular grid, there may not be enough data locations (there is no magic number for  how many are needed)

1987, A. Warrick and D.E. Myers, Optimization of Sampling Locations for Variogram Calculations Water Resources Research 23, 496-500

Finally your question suggests that you need to consult a book on geostatistics to get a better understanding, see for example "Geostatistics" by J.P. Chiles and P. Delfiner, 2nd edition, John Wiley & Sons

1990, A.W. Warrick, R. Zhang, M.M. Moody and D.E. Myers, Kriging Versus Alternative Interpolators: Errors and Sensitivity to Model Inputs - in Field-Scale Water and Solute Flux in Soils (Monte Verita),(eds)Roth, Fl�hler, and Jury � Birkhauser Verlag Basel