| Efficient sampling and effective use of data are desirable because the number of experiments that might be undertaken is limited by time and cost. The precision of prediction depends on the number and precision of samples, which have a trade-off relation in sampling. Whereas low precision data requires a lot of samples to improve prediction precision, obtaining high precision samples increases the required time and cost. We propose a prediction and optimization method which combines different samples of varying precision. Data is divided into several levels according to its precision and then an auto-regression model with a Gaussian process is assumed between each level. On every level the prediction is updated by adding higher precision samples. Then, we consider the use of Efficient Global Optimization (EGO) as an indicator of the possibility of a solution being optimal. This indicator consists of the predicted value and the accuracy of the predictor at each point. The prediction is sequentially updated by adding higher precision samples which are selected by the indicator. By updating the prediction, the optimal solution can be searched for with an efficient sampling method. Finally, the precision and efficiency of optimization is examined through numerical simulations which utilize the proposed method. |
|
