Improved decision support for engine-in-the-loop experimental design optimization

Abstract: Experimental optimization with hardware in the loop is a common procedure in engineering and has been the subject of intense development, particularly when it is applied to relatively complex combinatorial systems that are not completely understood, or where accurate modelling is not possible owing to the dimensions of the search space. A common source of difficulty arises because of the level of noise associated with experimental measurements, a combination of limited instrument precision, and extraneous factors. When a series of experiments is conducted to search for a combination of input parameters that results in a minimum or maximum response, under the imposition of noise, the underlying shape of the function being optimized can become very difficult to discern or even lost. A common methodology to support experimental search for optimal or suboptimal values is to use one of the many gradient descent methods. However, even sophisticated and proven methodologies, such as simulated annealing, can be significantly challenged in the presence of noise, since approximating the gradient at any point becomes highly unreliable. Often, experiments are accepted as a result of random noise which should be rejected, and vice versa. This is also true for other sampling techniques, including tabu and evolutionary algorithms