Knowledge Learning-Based Dimensionality Reduction for Solving Large-Scale Sparse Multiobjective Optimization...

Area of Science:

• Optimization
• Computational Intelligence
• Machine Learning

Background:

• Large-scale sparse multiobjective optimization problems (LSMOPs) are crucial in fields like critical node detection and feature selection.
• Existing evolutionary algorithms struggle with LSMOPs due to large search spaces and fixed dimensionality reduction, leading to local optima.
• Efficiently approximating sparse Pareto optimal solutions in LSMOPs remains a significant challenge.

Purpose of the Study:

• To develop an adaptive dimensionality reduction approach for LSMOPs that overcomes the limitations of fixed schemes.
• To improve the efficiency and adaptability of evolutionary algorithms in solving LSMOPs.
• To balance exploration and exploitation effectively during the evolutionary process.

Main Methods:

• A knowledge learning-based dimensionality reduction strategy is proposed.
• The impact of various dimensionality reduction schemes is evaluated early in the evolution.
• A multilayer perceptron learns from the evolutionary process to map features to optimal reduction schemes, recommending the best one each generation.

Main Results:

• The proposed approach demonstrates superior performance compared to existing evolutionary algorithms on benchmark and real-world LSMOPs.
• Adaptive dimensionality reduction effectively balances exploration and exploitation.
• Improved approximation of sparse Pareto optimal solutions was achieved.

Conclusions:

• The knowledge learning-based dimensionality reduction approach significantly enhances evolutionary algorithms for LSMOPs.
• Adaptive scheme selection leads to more robust and efficient optimization.
• This method offers a promising direction for tackling complex, large-scale optimization challenges.