A novel approach to model design and tuning, through automatic parameter screening and optimization

摘要

The aim of this paper is to describe a novel methodology that has been developed and tested for improving the process of model tuning, essential step for the certification of full flight simulators. We address Tuning because, at the state-of-the-art, the development of life-critical simulations requires months to appropriately tune the model. The proposed approach is based on automatic techniques for parameter screening, i.e. identification of the relevant parameters to simulate a system, and optimization, i.e. search of optimal values for those parameters. All techniques are fully general, because they leverage ideas from Machine-Learning and Optimization Theory to achieve their goals without directly analysing the simulator's mathematical model. Concerning screening, we show how simple Machine-Learning algorithms, based on Neural Networks and Logistic Regression, can be used for ranking the parameters according to their relevance. Concerning optimization, we describe two algorithms: an adaptive hill-climbing procedure and a novel strategy, which we called sequential masking. Eventually, we show the performances achieved and the impact on the time and effort required for tuning a helicopter flight-simulator.