Combining statistical models

Suppose you have n learned statistical models and you want to combine them to archive best score possible.

At first I combined like this:

Out = Sum(p[i]*out[i]);
 where p[i] - normalized manually tuned coefficients

I tried to analyze where each model behaves better and manually tune coefficients. Then I just trained individual out tuples to regression tree and got huge improvement, 
that I've never archived manually