Multigranularity Surrogate Modeling for Evolutionary Multiobjective Optimization With Expensive Constraints.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning Applications

Background:

• Multiobjective optimization problems (MOPs) with expensive constraints are computationally challenging for surrogate-assisted evolutionary algorithms (SAEAs).
• Existing SAEAs struggle with approximating constraint violation (CV) or individual constraints due to complexity, cumulative errors, and high costs.
• A single approximation granularity limits the effectiveness of SAEAs in handling diverse constraint landscapes.

Purpose of the Study:

• To develop a novel multigranularity surrogate modeling framework for evolutionary algorithms (EAs) to address challenges in MOPs with expensive constraints.
• To adaptively determine the approximation granularity of constraint surrogates based on population position in the fitness landscape.
• To introduce a model management strategy to mitigate errors and prevent local optima entrapment.

Main Methods:

• A multigranularity surrogate modeling framework is proposed, adapting approximation granularity for constraint surrogates.
• A dedicated model management strategy is incorporated to reduce surrogate errors and enhance exploration.
• An implementation, K-MGSAEA, is developed to evaluate the proposed framework's performance.

Main Results:

• The proposed K-MGSAEA framework demonstrates superior performance compared to seven state-of-the-art competitors.
• Experimental results on numerous test problems validate the effectiveness of the multigranularity approach.
• Adaptive granularity and model management significantly improve handling of expensive constraints in MOPs.

Conclusions:

• The developed multigranularity surrogate modeling framework offers a significant advancement for SAEAs tackling MOPs with expensive constraints.
• The adaptive strategy effectively balances approximation accuracy and computational cost.
• K-MGSAEA provides a robust and efficient solution, outperforming existing methods in challenging optimization scenarios.