Dynamic multi-objective evolutionary optimization algorithm based on two-stage prediction strategy.

Area of Science:

• Computational Intelligence
• Evolutionary Computation
• Optimization Algorithms

Background:

• Dynamic multi-objective optimization evolutionary algorithms (DMOEAs) face challenges in tracking Pareto-optimal sets efficiently.
• Existing DMOEAs exhibit deficiencies such as random search in early stages and underutilization of knowledge for convergence acceleration in later stages.

Purpose of the Study:

• To propose a novel DMOEA, the two-stage prediction strategy (TSPS), to address the limitations of current algorithms.
• To enhance the convergence rate, population diversity, and responsiveness to environmental changes in DMOEAs.

Main Methods:

• The proposed TSPS divides the optimization process into two stages: multi-region knee point selection for Pareto-optimal front shape capture and improved inverse modeling for searching representative individuals.
• Stage one focuses on accelerating convergence and maintaining diversity by identifying key points on the Pareto front.
• Stage two enhances population diversity and aids in predicting the Pareto front's movement.

Main Results:

• Experimental results demonstrate that TSPS significantly outperforms six other DMOEAs on standard test suites.
• The proposed method exhibits a superior ability to respond rapidly to environmental changes in dynamic optimization problems.
• TSPS effectively balances convergence speed and population diversity throughout the optimization process.

Conclusions:

• The two-stage prediction strategy (TSPS) offers a robust and efficient approach for dynamic multi-objective optimization.
• TSPS provides a significant advancement over existing DMOEAs, particularly in complex and rapidly changing environments.
• The method's effectiveness in accelerating convergence, maintaining diversity, and adapting to environmental shifts validates its utility.