It is appropriate to distinguish among the three steps of the approach indicated in Fig. 1. The most bulky part concerns the data base generation and the pre-computation of the LPMs. For example, the computation of the 135,000 LPMs of our data base took all in all about 1 month CPU on a SUN Sparc10 workstation (18MFLOPS). However, using several high-end workstations in parallel, the elapsed time may be easily reduced to less than one day. The second step, consisting of extracting the regression models from a data base, is much faster, although it remains off-line task. To fix ideas, on the same hardware it would typically take some minutes to build a regression tree and some hours to optimize the weights of a multilayer perceptron. Finally, the on-line use of the trees or the perceptrons is extremely fast, and would typically take less than a millisecond per margin estimation.