A Global Constrained Optimization Algorithm for Engine Calibration

Abstract

The engine calibration procedure developed in Renault involves solving a constrained global optimization problem. This is a medium scale problem with a great number of linear and non linear constraints and containing a large number of local minima. Therefore, we use a strategy combining global and local optimization. We propose a new global stochastic optimization algorithm called Multistoch generating a fixed number of starting points for a local optimization procedure (see 2). These set of points are called a grid. This algorithm although different from simulated annealing (see 3) involves a gibbs measure. This measure is used to select a point in the grid. Around this selected point we generate a new candidate point which according to his value will modify the grid. After a description of the engine calibration problem some theorical and numerical results on the algorithm are presented.