Hierarchical Stochastic Optimization With Application to Parameter Tuning for Electronically Controlled Transmissions

In mechanical systems, control parameters are often manually tuned by an expert through trial and error, which is labor-intensive and time-consuming. In addition, the difficulty of this problem is that there often exist multiple solutions that provide high returns. As a designed objective function is often not optimal in practice, the solution that provides the highest return may not be the optimal solution. Therefore, it is often necessary to verify the multiple candidates of the solution to identify the one most suitable for the actual system. To address this issue, we propose a parameter optimization system using hierarchical stochastic optimization (HSO) that can handle multimodal objective functions. In a case study of electronically controlled transmissions, the optimizer learns multiple sets of parameters that satisfy all constraints and outperforms the parameters manually designed by human engineers. We demonstrate experimentally that our HSO can identify several modes of the objective function and is more sample-efficient than the existing methods, such as cross-entropy method and covariance matrix adaptation evolution strategy, as well as a human engineer.