I don't really understand the question. Regression is done the same way regardless of how many predictors there are; what exactly is the problem that you are having?
In the presence of many variables the standand error of the lineal regression equations decreases, between other reasons, because the best ajustment model securely is non lineal, mainly for the relationships related to engineering.
I have worked the non-lineal regression equations included posinomials and signomials. Your model could be complex, including different non lineal terms.
On the other hand, your work to do with an engineering task for what I could to recommend you a methodology that you can to find in the work "Analysis and Synthesis of Engineering Systems" that you can easily find in my reseachgate profil.
I am well aware of regression analysis technique. Actually, I wanted a research paper published which contains regression equation of large number of variables that I could use for optimisation.
It's an interesting question. What about overfitting? For a large number of predictors, you need a large number of observations. I haven't seen the use of such numbers (15) in engineering. What could be the influence of the 15th predictor with the lowest influence on the output? Why not working on several models with a reduced number of predictors to identify the key predictors and then build a final model for your problem?
For using an optimization procedure you need to obtain relations with the maximum possible precision, helped by regression non lineql analysis or by neural artificial networks. Of course your relationship will always have some error of determination, for what you could to optimize the main value of the efficiency indicator, it value assured with certain probability, etc.