Meta-Learning-Based Deep Reinforcement Learning for Multiobjective Optimization Problems.

Area of Science:

• Artificial Intelligence
• Operations Research
• Computer Science

Background:

• Deep reinforcement learning (DRL) shows promise for combinatorial optimization.
• Existing DRL methods struggle with multiobjective problems due to complex subproblem decomposition.
• Efficiently handling multiple objectives remains a challenge in DRL applications.

Purpose of the Study:

• To propose a concise meta-learning-based DRL approach for multiobjective optimization.
• To address the limitations of existing DRL methods in flexibly and efficiently dealing with multiple subproblems.
• To improve the quality and diversity of solutions for complex multiobjective problems.

Main Methods:

• A meta-model is trained using meta-learning.
• The meta-model is fine-tuned to derive submodels for specific subproblems.
• Pareto fronts are constructed based on the derived submodels.

Main Results:

• The proposed method significantly reduces training time for multiple submodels compared to other learning-based methods.
• The meta-model's adaptability allows for derivation of more submodels, enhancing solution quality and diversity.
• Experimental results on multiobjective traveling salesman and vehicle routing problems demonstrate superiority over existing approaches.

Conclusions:

• The meta-learning-based DRL approach offers a superior solution for multiobjective combinatorial optimization.
• This method provides rapid adaptability and improved efficiency in deriving submodels.
• It enhances the quality and diversity of solutions for complex optimization tasks.