Development and Application of a Non-Gradient Step-Controlled Search Algorithm for Engine Combustion Optimization 2006-01-0239
A new search technique, called Non-Gradient Step-Controlled algorithm (NGSC), is presented. The NGSC was applied independently from pre-selected starting points and as a supplement to a Genetic Algorithm (GA) to optimize a HSDI diesel engine using split injection strategies. It is shown that the NGSC handles well the challenges of a complex response surface and factor high-dimensionality, which demonstrates its capability as an efficient and accurate tool to seek “local” convergence on complex surfaces. By directly tracking the change of a merit function, the NGSC places no requirement on response surface continuity / differentiability, and hence is more robust than gradient-dependent search techniques. The directional search mechanism takes factor interactions into consideration, and search step size control is adopted to facilitate search efficiency.
The NGSC was also coupled with a Random Number Generator (RNG) to formulate a global optimization framework called a Path-Tracking Global NGSC (PTGN) methodology. This was supplemented with additional features such as the use of a path-tracking mechanism, a time-varying convergence threshold, and a Gray initialization code to improve the overall search quality. By comparing to optimizations performed using GA, it is shown that the PTGN discovered various promising regions, some of which were not found by the GA, and in less evaluation attempts. Moreover, the path-tracking mechanism, combined with the excellent local convergence capability of the NGSC provides a methodology for global convergence.
As an application, a systematic engine optimization framework work was formulated, which combines tools of search, evaluation and data analysis. An advanced non-parametric regression method, the Component Selection and Smoothing Operator (COSSO), was applied to Computational Fluid Dynamics code data drawn from the optimizations. Models were constructed to correlate the response (merit) and control factors. In this way, the sensitivity of responses to each control factor could be quantitatively interpreted, and this was used to judge the “goodness” of predicted optima.