Engine control is based today on open-loop control structures which force the combustion input quantities to follow references chosen heuristically to ensure a sufficiently clean and efficient combustion. The references are usually produced by some simple algorithm tuned with the real system including the actuators, and thus keeping in account their limits. While this approach does offer operating simplicity and reliability, it is clearly hardly adequate for transient conditions and proves therefore suboptimal. In this paper, we propose to use structure identification and numerical optimization tools based on input/output models of the emissions to compute the ideal references. This approach requires a very limited measurement time and some off-line computations, being very simple to implement. This paper applies it for a 4 cylinder Diesel engine under the action of a multi-linear predictive control and shows that large improvements of total emissions are possible, but are also strictly related to the dynamic performance of the actuators.