A surrogate-ensemble assisted expensive many-objective optimization

Yi Zhao,Chaoli Sun,Jianchao Zeng,Ying Tan,Guochen Zhang

DOI: https://doi.org/10.1016/j.knosys.2020.106520

2021-01-01

Abstract:<p>The surrogate ensemble has been shown to have a good performance for assisting the evolutionary algorithms to solve computationally expensive single-objective problems. However, it has not been paid much attention to assist in multi-/many-objective optimization. In this paper, we propose to train multiple surrogate models to assist many-objective optimization algorithm for solving expensive many-objective problems. In the proposed method, some non-dominated solutions, which have been evaluated using the expensive objectives, will be utilized together with the current population to generate the offspring in order to prevent the population from moving forward to a bad region. In the model management, the way to determine the method used for selecting individuals to be evaluated using the exact expensive objectives depends on the convergence degree of the current population. A new method to measure the uncertainty is also proposed in our method by considering the distance to the samples in the decision space and the approximation variance in the objective space. The experimental results on a number of benchmark problems show that our proposed method is competitive to two state-of-the-art surrogate-assisted evolutionary algorithms for solving expensive many-objective problems in a limited computational budget.</p>

computer science, artificial intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to effectively carry out multi - objective optimization when the computational cost is high, especially when the number of objectives is large (i.e., multi - objective or many - objective optimization problems). Specifically, the paper proposes a surrogate - ensemble - assisted method to solve expensive multi - objective optimization problems, aiming to reduce the number of actual objective function evaluations while maintaining or improving the performance of the optimization algorithm.

### Core Problems of the Paper
1. **Expensive Multi - objective Optimization Problems**: In many practical applications, the evaluation of each objective function can be very time - consuming and resource - intensive, especially when the number of objectives is large. Therefore, how to find a solution close to the optimal solution within a limited computational budget is a challenge.
2. **Selection and Management of Surrogate Models**: Most of the existing surrogate - assisted evolutionary algorithms mainly focus on single - objective optimization problems, and there is less research on multi - objective optimization, especially many - objective optimization problems. In addition, a single type of surrogate model may not be suitable for all types of problems, so a method for selecting and managing multiple surrogate models is required.

### Solutions
The paper proposes a method named SAEMO (Surrogate - ensemble Assisted Expensive Many - objective Optimization), and its main contributions are as follows:

1. **Retention of Non - dominated Solutions**: In order to avoid the error accumulation caused by the use of approximation techniques, the paper proposes to retain some non - dominated solutions that have been evaluated by the real objective function in each generation and use them together with the current population to generate offspring.
2. **New Uncertainty Measurement Method**: A new method for measuring the uncertainty of surrogate models is proposed, which takes into account the sample distance in the decision space and the approximate variance in the objective space.
3. **Convergence Quality Measurement**: A method for measuring the convergence quality of the population is proposed and used as a criterion for selecting individuals for real objective function evaluation to balance exploration and exploitation.

### Method Overview
- **Initialization**: Generate an initial population and evaluate it using the real objective function, save all individuals to the archive \( D_{\text{archive}} \), and save the non - dominated individuals to another archive \( ND_{\text{archive}} \).
- **Generate Mating Pool**: Select some non - dominated solutions from \( ND_{\text{archive}} \) and generate offspring together with the current population.
- **Environmental Selection**: Use the surrogate - ensemble model to approximately evaluate each individual in the combined population and select the next - generation population according to the approximate values.
- **Individual Selection**: Select individuals for real objective function evaluation according to the convergence quality and uncertainty measurement of the population.

### Formula Summary
- **Objective Vector Transformation**:
\[
\mathbf{F}'(\mathbf{x})=\mathbf{F}(\mathbf{x})-\mathbf{Z}^*
\]
where \(\mathbf{F}(\mathbf{x})=(f_1(\mathbf{x}), f_2(\mathbf{x}), \ldots, f_m(\mathbf{x}))\) is the objective vector, and \(\mathbf{Z}^*\) is the point composed of the minimum values of each objective in \(ND_{\text{archive}}\).
- **Angle Calculation**:
\[
\theta_{i,j}=\arccos\left(\frac{\mathbf{F}'(\mathbf{x}_i)\cdot\mathbf{v}_j}{\|\mathbf{F}'(\mathbf{x}_i)\|}\right)
\]
where \(\mathbf{v}_j\) is the reference vector.
- **Output of Surrogate - ensemble Model**:
\[
\tilde{f}_i(\mathbf{x})=\frac{\tilde{f}_{PR,i}(\mathbf{x})+ \