Hello Vincenzo,
one option that would be easy to try is to calculate the residuals (y - f(x)) from the first pass and then use the residuals as the target. This way you could try to build up an additive model y = f1(x) + f2(x) + ... . This is essentially the boosting idea (see e.g. [1])
The GP literature describes some ideas how a population might be seeded using information gathered from earlier runs. However, to test such ideas in HeuristicLab you would need to look into implementing your own tree creation / population initialization routine.
Best, Gabriel