It is pretty straightforward.  Let's say you have two variables, a and b,and four different models with the following corresponding weights:

Model     Weight

nul            0.3

a               0.3

b               0.05

a,b            0.35

Notice the weights sum to 1 as they should.  All you do for each variable is sum the weights of the models where it is included.  So for variable a it would equal 0.3 + 0.35 = 0.65 and for b it would equal 0.05 + 0.35 = 0.4,  From this you would conclude that a is a more important in explaining the variation in the data than  b is.  The number of models and the number of variable does not matter.  HOWEVER, it is critical that each variable be present in the same number of models.  In this case both a and b are included in 2 models.  This method does not work if this is not the case.  For example if you included an interaction term a*b you might be tempted to fit only the following models:

Model

nul

a

b

a,b

a,b,a*b

However, because a and b are included in in 3 models each and a*b in only one you could not calculate the relative importance of each variable.  In order to do this you would have to fit all the following models (which is generally possible):

Model

nul

a

b

a*b

a,b

a,a*b

b,a*b

a,b,a*b

In this case, each variable is in 4 models so you could sum the weights for each variable.

Citation for this would be the Burnham and Anderson book:

Burnham, K. P., and Anderson, D. R. 2002. Model selection and multimodel inference: A practical information-theoretic approach, Springer Science + Business Media, New York.

I hope that helps.

Thanks so much William,

I was referring if there is any standard values to make a decission about the power of the variables. I mean, if the relative importance is 0.7, the effect is strong, if is 0.5 the effect is moderate, etc.

In short, no.  This is because the number of models and the number of variables can differ.  In the first case I show above, you have two variables that are each in 2 of 4 models.  If you had instead a series of 6 variables that were each in 2 of 12 models, then the average weight for each variable is going to be a lot lower because the overall sum of the weights is always going to be 1.  This measure simply tells you the relative importance of each variable to the other variables in the set. 

You can follow that Kittle et al. 2008 says about calculate the relative importance of predictors in a unbalanced models set, where them recurre to ponderate the cumulative of wi for each predictor divided for the total numbers of models were each predictor is represented.

textual, them says: " To determine the relative importance of variables and further reduce model ambiguity, we summed wi of all models containing a common factor for each analysis. The higher the combined weights for an explanatory variable, the more important it is for the analysis (Burnham and Anderson 2002). For this measure to be meaningful it is necessary to have the same number of models containing each variable (Burnham and Anderson 2002), so we divided the cumulative model weights for a particular variable by the number of models containing that variable to get an average variable weight (wi) per model "

DOI 10.1007/s00442-008-1051-9