Heterogeneous Ensemble-Based Infill Criterion for Evolutionary Multiobjective Optimization of Expensive Problems.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Gaussian processes (GPs) are widely used in surrogate-assisted evolutionary optimization due to their uncertainty estimation capabilities.
• However, the computational cost of GPs increases significantly with more training data, limiting their scalability.
• This poses a challenge for optimizing computationally expensive problems using GPs.

Purpose of the Study:

• To propose and evaluate heterogeneous ensembles as a scalable surrogate modeling approach for evolutionary optimization.
• To enhance ensemble reliability for uncertainty estimation and promote diversity in surrogate models.
• To compare the performance and scalability of heterogeneous ensembles against Gaussian processes and homogeneous ensembles.

Main Methods:

• Constructed a heterogeneous ensemble using a least square support vector machine and two radial basis function networks.
• Incorporated original, subset, and transformed decision variables as inputs to the ensemble.
• Evaluated the heterogeneous ensemble against Gaussian processes and homogeneous ensembles in evolutionary multiobjective optimization.

Main Results:

• The heterogeneous ensemble demonstrated competitive performance compared to Gaussian processes.
• The proposed ensemble showed significantly better scalability in computational complexity with increasing search dimensions.
• The use of diverse inputs enhanced the ensemble's ability to promote search diversity.

Conclusions:

• Heterogeneous ensembles offer a viable and scalable alternative to Gaussian processes for surrogate-assisted evolutionary optimization.
• The proposed ensemble approach effectively balances performance and computational efficiency.
• This method is particularly beneficial for computationally expensive optimization problems where GP scalability is a concern.