Learning to Optimize: Reference Vector Reinforcement Learning Adaption to Constrained Many-Objective Optimization of...

Area of Science:

• Optimization algorithms
• Machine learning applications
• Industrial engineering

Background:

• Decomposition-based algorithms' performance is limited by fixed reference vectors.
• Lack of adaptability to diverse problem characteristics hinders optimization.
• Industrial copper burdening presents complex optimization challenges.

Purpose of the Study:

• To propose an adaptive reference vector reinforcement learning (RVRL) approach.
• To enhance the adaptability of decomposition-based algorithms for industrial optimization.
• To address the challenges of industrial copper burdening optimization.

Main Methods:

• A reinforcement learning (RL) operation for adaptive reference vector adjustment.
• A reference point sampling operation using estimation-of-distribution learning.
• Integration of adaptive penalty functions and soft constraint relaxation for complex constraints.

Main Results:

• The RVRL approach demonstrates improved adaptability to problem characteristics.
• The algorithm effectively handles complex constraints in industrial copper burdening.
• Experimental results confirm the competitiveness and effectiveness on benchmarks and real-world instances.

Conclusions:

• The proposed RVRL approach offers a significant advancement for decomposition-based optimization.
• This method provides a robust solution for complex industrial optimization tasks.
• The study validates the efficacy of adaptive strategies in tackling real-world engineering problems.