An Adaptive Reference Vector-Guided Evolutionary Algorithm Using Growing Neural Gas for Many-Objective Optimization...

Area of Science:

• Computational Intelligence
• Multiobjective Optimization
• Artificial Neural Networks

Background:

• Reference vector-based decomposition algorithms struggle with irregular Pareto fronts (PFs) due to poor vector distribution.
• Existing adaptive methods for reference vectors can slow down convergence by frequently updating vectors based on activity.
• Efficiently handling complex and irregular PFs is crucial for advancing multiobjective optimization.

Purpose of the Study:

• To develop a novel method for automatic and stable adaptation of reference vectors in decomposition-based multiobjective optimization.
• To address the limitations of existing adaptive techniques that hinder search process convergence.
• To improve the performance of algorithms dealing with irregular Pareto fronts.

Main Methods:

• Proposed a new method employing a growing neural gas (GNG) network to learn the distribution of reference vectors.
• Designed an improved GNG for learning the topology of PFs using solutions from the search process as training data.
• Utilized individuals from current and previous generations to train the GNG, balancing exploration and exploitation.

Main Results:

• The proposed GNG-based method achieves automatic and stable adaptation of reference vectors.
• Demonstrated competitive performance on popular benchmark problems with complex and irregular PFs.
• Showcased effectiveness on a real-world hybrid vehicle controller design problem.

Conclusions:

• The GNG-based approach offers a robust solution for adapting reference vectors in multiobjective optimization with irregular PFs.
• The method effectively balances exploration and exploitation, leading to improved search efficiency.
• This technique presents a promising advancement for tackling challenging multiobjective optimization problems.