Does anyone know of a function approximator which can produce a variable number of output values (i.e. for some regions in input space it might output a vector of 3 values, whereas in other regions it might produce 5 outputs)?

Update: Thanks everyone for your suggestions. I realise now that I missed a critical aspect when phrasing my original question - we don't know in advance how many outputs will be required in each region of the input space (or even what the regions of the input space are). So maybe, I should rephrase my question in light of Simone and Meysar's answers - is there a function approximator which can learn to produce a single output for some parts of input space, and no output for other parts? My thinking so far is to use something like an RBF network as suggested by Vassilis, with a threshold applied so no output is produced if the input doesn't closely match any of the basis functions.

Similar questions and discussions