Automated Configuration of Evolutionary Algorithms via Deep Reinforcement Learning for Constrained Multiobjective...

Area of Science:

• Evolutionary Computation
• Artificial Intelligence
• Optimization Algorithms

Background:

• Automated algorithm design is crucial for constrained multiobjective optimization evolutionary algorithms (CMOEAs).
• Current learning-assisted CMOEAs rely on manual, often suboptimal, expert-designed techniques.
• Existing methods lack versatility and adaptability in dynamic optimization landscapes.

Purpose of the Study:

• To develop a versatile and effective automated configuration method for CMOEAs.
• To leverage deep reinforcement learning (DRL) for self-adapting CMOEA parameters and operators.
• To enhance the performance and adaptability of CMOEAs through automated design.

Main Methods:

• Transformed CMOEA online configuration into discrete and continuous parameter determination.
• Applied deep reinforcement learning (DRL), specifically Actor-Critic and deep Q-learning, for automated configuration.
• Developed a novel CMOEA incorporating the automatically configured evolutionary algorithm (EA).

Main Results:

• The DRL-configured CMOEA demonstrated significant performance improvements over 11 state-of-the-art methods.
• Experiments on challenging benchmarks and real-world problems validated the proposed method's superiority.
• The automated configuration showed greater versatility and effectiveness compared to handcrafted approaches.

Conclusions:

• Automated configuration using DRL offers a promising direction for advancing evolutionary multiobjective optimization.
• The self-learning capability of the DRL-configured CMOEA enhances its adaptability and performance.
• This work establishes a new paradigm for designing versatile and high-performing CMOEAs.