Abstract:In engineering, models are often used to represent the behavior of a system. Estimators are then needed to approximate the values of the model's parameters based on observations. This approximation implies a difference between the values predicted by the model and the observations that have been made. It creates an uncertainty that can lead to dangerous decision making. Interval analysis tools can be used to guarantee some properties of an estimator, even when the estimator itself doesn't rely on interval analysis (Adam, 2019) (Adam, 2015). This paper contributes to this dynamic by proposing an interval-based and guaranteed method to validate a nonlinear estimator. It is based on the Moore-Skelboe algorithm (van Emden, 2004). This method returns a guaranteed maximum error that the estimator will never exceed. We will show that we can guarantee properties even when working with non-guaranteed estimators such as neural networks.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to verify the maximum error of non - linear estimators and ensure their reliability in specific applications. Specifically, the authors propose a method based on interval analysis to verify the performance of non - linear estimators, especially for those estimators that do not provide guaranteed results by themselves (such as neural networks). Through this method, an upper bound of the maximum error can be obtained, thus providing a reliable basis for decision - making.
### Problem Background
In the engineering field, models are usually used to represent the operating behavior of systems. In order to make the model better fit the actual system, estimators are required to approximate the values of model parameters according to the observed data. However, this approximation will inevitably produce errors, and these errors may lead to dangerous decisions. Therefore, it is very important to accurately characterize and verify the errors of estimators.
For the linear parameter estimation problem (LPE), there are already many mature methods (such as the least - squares method) that are efficient and easy to use. But for the nonlinear parameter estimation problem (NLPE), the existing methods are often less efficient and cannot provide strict guarantees. To overcome this problem, estimators based on interval analysis have been developed, but their applications depend on the understanding of interval analysis tools, which is not always applicable in some application scenarios.
### Paper Contributions
The main contribution of this paper is to propose an interval analysis method based on the Moore - Skelboe algorithm to verify the maximum error of non - linear estimators. This method can be applied to any non - linear estimator, including neural networks, and provides a strict upper bound of the maximum error. Even if the estimator itself does not provide guaranteed results, this method can also ensure the upper limit of the error.
### Method Overview
1. **Problem Definition**:
- Define the estimator \(\psi\) and the error function \(\epsilon(x, e)\), where \(x\) is the true parameter and \(e\) is the noise.
- The goal is to find the maximum value \(\bar{\epsilon}\) of the error function.
2. **Interval Analysis and Boxes**:
- Use intervals to represent the ranges of parameters and noise.
- Use the Cartesian product of intervals to form multi - dimensional boxes.
3. **Optimization Problem**:
- Transform the error maximization problem into an optimization problem and solve it using the Moore - Skelboe algorithm.
4. **Moore - Skelboe Algorithm**:
- Continuously divide and evaluate boxes to gradually narrow the value range of the objective function and finally obtain the interval containing the global minimum.
### Application Example
The paper shows the application of this method through a specific example. Suppose we want to estimate a two - dimensional position \(x\), and the observed data are the distances from \(x\) to three known landmarks \(a, b, c\). By training a neural network as an estimator and using the Moore - Skelboe algorithm to verify its maximum error, the result shows that the maximum error is \(\bar{\epsilon}=1.7\).
### Conclusion
The interval - analysis - based method proposed in this paper can effectively verify the maximum error of non - linear estimators, and can provide a reliable error upper bound even for estimators that do not provide guaranteed results by themselves, such as neural networks. This is of great significance for improving the reliability of non - linear estimators in practical applications.