An Analysis of AIC for Linear Stochastic Regression and Control

Stimulated by the industrial successes of Akaike´s minimum FPEC criterion (MFPEC) for selecting input variables for LQG controllers, we have sought a theoretical statistical framework which would help to explain its utility as a criterion for comparing diverse models. It turns out that Akaike´s Entropy Maximization Principle can be shown to provide such a framework. This paper summarizes some new results concerning this principle and linear stochastic regression version of Akaike´s minimum AIC criterion (MAIC) which is asymptotically equivalent to MEPEC. Some analyses related to a successful ship autopilot design project are presented to illustrate the application and scope of MAIC.