Slide 30: In conclusion
In general, it has been shown that neural networks (which have the ability to "learn" the important variables in a system) and GAMs (possibly with an autocorrelation coefficient) are the most robust methods using minimal data. The problem with neural networks is that they are "black boxes" - that is, the decision processes within the networks are generally unavailable for examination, making it a challenge to derive which are the governing determinant factors.
Thus, there are two real options for deciding on the best model - firstly, to use an expert system such as BIOMOD, to try several different models and derive the one that best decribes the observed data, or to use a model that is generally robust. Secondly, a model can be selected that is specifically suited to the questions being asked, such as Ecological Niche Factor Analysis when there is no absence data to use.
However, generalized additive models with pseudo-absence are generally robust, and may often out-perform presence-only techniques (Brotons et al 2004).