Page 248 -

236 5 Neural Networks

MLP1:3:1 solution using variable GIR is found, with higher correlation and
smaller errors, which is shown in Figure 5.60a. By also performing a medium IPS
search in the feature space without variable GIR, it was possible to find an
alternative MLP3:2:1 solution using variables CA, DEPR and AIC, with much
poor correlation (around 0.3). This solution took about 28 seconds to find on the
same 733 MHz Pentium. The respective regression result is depicted in Figure
5.60b.

Figure 5.60. Regression (grey line) of the capital revenue (CapR) for 500
industrial firms, sorted (Case) by increasing values of CapR (black line). (a) Using
an MLPI:3:1 net with variable GIR as input (correlation = 0.8) ; (b) Using an
MLP3:2:1 net with variables CA, DEPR and A/C as inputs (correlation = 0.3).