Two-Stage Data-Driven Evolutionary Optimization for High-Dimensional Expensive Problems.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Surrogate-assisted evolutionary algorithms (SAEAs) are crucial for expensive optimization problems.
• Existing SAEAs often suffer from slow convergence in later optimization stages.

Purpose of the Study:

• To propose a novel two-stage data-driven evolutionary optimization (TS-DDEO) framework.
• To address the limitations of slow convergence in existing SAEAs, particularly in the exploitation phase.

Main Methods:

• A two-stage approach combining hierarchical particle swarm optimization for exploration and best-data-driven optimization (BDDO) for exploitation.
• BDDO features real-time surrogate model and population updates using top-ranked solutions.
• Incorporation of three surrogate-assisted sampling strategies: differential evolution, local search, and a novel full-crossover (FC) strategy.

Main Results:

• Experimental validation confirms the effectiveness of the TS-DDEO framework, BDDO method, and FC strategy.
• The proposed TS-DDEO algorithm demonstrates superior performance compared to five state-of-the-art SAEAs on high-dimensional benchmark functions.
• Demonstrated improvements in both solution effectiveness and algorithmic robustness.

Conclusions:

• The TS-DDEO framework effectively balances exploration and exploitation for computationally expensive optimization.
• The BDDO method significantly accelerates the optimization process with its efficient update mechanisms.
• The novel FC strategy contributes to integrating valuable genetic information for enhanced performance.