Expensive Multiobjective Optimization Guided by Attention-Enhanced Generative Models.

Area of Science:

• Optimization
• Artificial Intelligence
• Computational Science

Background:

• Expensive multiobjective optimization problems require efficient algorithms.
• Existing surrogate-assisted evolutionary algorithms (SAEAs) often use inefficient genetic operators.
• There is a need for advanced methods to generate high-quality solutions in complex optimization tasks.

Purpose of the Study:

• To develop a novel learning-based generative model to replace conventional genetic operators in SAEAs.
• To enhance the efficiency and effectiveness of multiobjective search for expensive problems.
• To introduce an attention-enhanced convolutional residual network for offspring generation.

Main Methods:

• A learning-based generative model (LMOGM) is proposed, utilizing an attention-enhanced convolutional residual network.
• The model generates promising solutions for subproblems using the Tchebycheff metric.
• A surrogate model is employed for online optimization of the generative model's hyperparameters.

Main Results:

• The proposed LMOGM demonstrates superior performance on DTLZ, ZDT, and WFG benchmark suites across various dimensions.
• The approach significantly outperforms traditional evolutionary algorithms and state-of-the-art SAEAs.
• Effective application in geothermal energy extraction design optimization is shown.

Conclusions:

• The learning-based generative model offers a more efficient approach to multiobjective optimization compared to existing SAEAs.
• This framework effectively addresses the limitations of conventional genetic operators in expensive optimization scenarios.
• The study highlights the potential of generative models in advancing the field of evolutionary computation.