Abstract:

The complex simulation contains many influencing factors, and its influence relationship is nonlinear, which need an efficient meta modeling method. The establishment of the meta model depends not only on the mathematical structure of the selected meta model, but also on the sufficiency of the simulation training data. Data adequacy directly affects the design of simulation experiment space and the computational complexity of experiment operation, so it is necessary to deal with the dimension explosion dilemma through targeted simulation experiment design methods. Sequential experimental design method can effectively improve the efficiency of design, but the traditional sequential method relies on human intervention and prior knowledge about the target system, which is not easy to carry out automatically. Therefore, a sequential neighbor exploration experiment design (SNEED) method is proposed to support simulation meta modeling. SNEED combines gradient sampling and random nearest neighbor exploration technology to generate as few experimental points as possible and extract information about experimental objects automatically, which can effectively improve the experimental efficiency of complex simulation meta modeling. In the case part, the proposed method is compared with Latin hypercube and traditional sequential sampling methods to verify the effectiveness of the proposed method.

Key words: simulation metamodeling, sequential experimental design, neighbor exploratory, Latin hypercube sampling