Adaptive dipole-like parameter calibration of complex black-box continuous processes

Large-scale systems such as flexible manufacturing, chemical production facilities, and traffic networks aim to maximize measures related to profit, health and safety, throughput, and service level. Due to the complexity of such systems, the mechanism that connects the input parameters to performance outputs is often unavailable, and optimization methods based on convexity or even differentiability of the objective function may not be applicable. Since these systems are characterized by heavy costs per unit time, the system manager has to resort to black-box approaches for optimization, where a set of parameters is tuned in order to maximize an accumulated performance measure of the process. In this paper, a novel mechanism is proposed for real-time calibration of parameters in continuous search spaces. The developed algorithm seeks the global optimum by means of solution exploitation, and adapts dynamically according to environmental changes. The solution method builds a sequence of random pairs of trials, called “dipoles”, which are used to adapt online the probability density function of the unknown parameters. The proposed method is characterized by the following advantages: (1) it does not depend on subjective coefficients setting; (2) solution exploitation starts from the first iteration; (3) the algorithm is effective also for systems with high dimensionality; (4) since sampling only involves two trials, exploitation is based on recent data rather than on data that extends far back in time. Several illustrative numerical examples are provided to show the applicability and efficiency of the proposed method.