Imitation Learning for Multiobjective Optimization-AlphaMOEA.

Area of Science:

• Artificial Intelligence
• Optimization
• Machine Learning

Background:

• Multiobjective optimization problems (MOPs) are challenging and have seen various multiobjective evolutionary algorithms (MOEAs).
• Existing MOEAs often require specific enhancements for different MOPs.

Purpose of the Study:

• Introduce AlphaMOEA, a novel artificial intelligence approach for solving MOPs.
• Demonstrate AlphaMOEA's effectiveness using an imitation learning-based end-to-end method.

Main Methods:

• AlphaMOEA employs a neural network architecture based on multitask learning (MTL).
• It involves two training stages: supervised learning (SL) to fit existing MOEA solutions and reinforcement learning (RL) for self-driven performance improvement.
• The RL stage utilizes similarity-based state design, an evolution operator-based action set, and an indicator-guided reward.

Main Results:

• AlphaMOEA effectively learns from high-dimensional representations of the decision space.
• The approach achieves a good balance between exploration and exploitation.
• Experimental results show improved performance in solving MOPs with diverse characteristics.

Conclusions:

• AlphaMOEA offers a new AI paradigm for MOPs, moving beyond traditional MOEAs.
• The model's ability to leverage high-dimensional knowledge enhances its problem-solving capabilities.
• AlphaMOEA demonstrates potential for efficient and effective MOP resolution.