Validation of neural networks in automotive engine calibration

This paper compares and contrasts different types of neural network error bars in the context of a real world safety critical problem, specifically the problem is one of the calibration of engine management systems for air to fuel ratio and ignition timing tables. Three types of error bars are considered and developed for radial basis function networks and Gaussian processes. The different assumptions inherent in the error bars are discussed in the context of a synthetic problem, and then applied to off-line and on-line engine data.