Abstract:This work is focused on finding G-optimal designs theoretically for kriging models with two-dimensional inputs and separable exponential covariance structures. For design comparison, the notion of evenness of two-dimensional grid designs is developed. The mathematical relationship between the design and the supremum of the mean squared prediction error ($SMSPE$) function is studied and then optimal designs are explored for both prospective and retrospective design scenarios. In the case of prospective designs, the new design is developed before the experiment is conducted and the regularly spaced grid is shown to be the G-optimal design. The retrospective designs are constructed by adding or deleting points from an already existing design. Deterministic algorithms are developed to find the best possible retrospective designs (which minimizes the $SMSPE$). It is found that a more evenly spread design under the G-optimality criterion leads to the best possible retrospective design. For all the cases of finding the optimal prospective designs and the best possible retrospective designs, both frequentist and Bayesian frameworks have been considered. The proposed methodology for finding retrospective designs is illustrated with a methane flux monitoring design.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to find theoretically optimal design schemes (G - optimal designs), especially for the Kriging model with two - dimensional input and separable exponential covariance structure. Specifically, the paper focuses on two aspects:
1. **Prospective Designs**: Determine the optimal grid design scheme before the start of the experiment. The paper proves that in the simple and ordinary Kriging models, the equispaced grid design is optimal, that is, it can minimize the supremum of the predicted mean squared prediction error (SMSPE).
2. **Retrospective Designs**: Improve the design by adding or deleting points on the basis of the existing design. The paper proposes a method based on the concept of "evenness" to compare different grid designs and develops deterministic algorithms for determining the best retrospective design. These algorithms can find the best design without relying on the values of covariance parameters.
### Main Contributions
1. **Theoretically prospective G - optimal grid design**: In the frequentist and Bayesian frameworks, the paper theoretically proves that the equispaced grid design is optimal.
2. **Grid evenness comparison criterion based on majorization**: Proposes a new method to compare the evenness of different grid designs and proves that a more even grid design can minimize SMSPE.
3. **Deterministic algorithms**: Provides algorithms for determining the best retrospective design, including cases of adding and deleting points.
### Mathematical Models and Symbols
- **Stochastic Process**: The paper considers a two - dimensional input stochastic process \(Z(x, y)\) whose covariance structure is a separable exponential covariance function:
\[
\text{Cov}(Z(x, y), Z(x', y'))=\sigma^{2} e^{-\alpha|x - x'|}e^{-\beta|y - y'|}
\]
where \(\sigma^{2}, \alpha, \beta>0\) are covariance parameters.
- **Mean Squared Prediction Error (MSPE)**: For a given design \(\xi=(d, \delta)\), the mean squared prediction error is defined as:
\[
\text{MSPE}((x_{0}, y_{0}), \xi, \Theta)=\text{Var}(Z(x_{0}, y_{0})-\hat{Z}(x_{0}, y_{0}))
\]
where \(\Theta = (\sigma^{2}, \alpha, \beta)\) is the parameter vector.
- **Supremum Mean Squared Prediction Error (SMSPE)**: Used to evaluate the overall performance of the design:
\[
\text{SMSPE}(\xi, \Theta)=\sup_{(x_{0}, y_{0})\in D_{1}\times D_{2}}\text{MSPE}((x_{0}, y_{0}), \xi, \Theta)
\]
### Algorithms and Results
- **Prospective Designs**: Proves that the equispaced grid design is optimal.
- **Retrospective Designs**:
- **Adding Points**: Proposes two methods: adding points one by one and adding points simultaneously. Through the concept of majorization, ensure that the new design is more even, thereby minimizing SMSPE.
- **Deleting Points**: Provides a deterministic algorithm to delete points from the existing design to find the best design.
### Case Application
The paper shows how to use the proposed retrospective design method to optimize the existing design through a case study of methane flux monitoring design.
In summary, through theoretical analysis and algorithm development, this paper systematically solves the optimal design problem of two - dimensional input in the Kriging model and provides strong support for experimental design in practical applications.