Reinforcement learning-controlled differential evolution with L-BFGS refinements

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Differential Evolution (DE) is a powerful optimization tool, but its performance relies heavily on parameter tuning.
• Existing adaptive DE strategies often use fixed rules and lack online learning, limiting their use in complex scenarios.
• Expensive black-box optimization requires adaptive methods that can learn from limited data.

Purpose of the Study:

• To develop a novel adaptive strategy for Differential Evolution using reinforcement learning.
• To create a data-driven framework for parameter adaptation in DE, moving beyond heuristic rules.
• To enhance DE's robustness and efficiency in high-dimensional and combinatorial optimization problems.

Main Methods:

• Proposed Reinforcement Learning-controlled Differential Evolution (RL-DE) with a three-layer closed-loop framework.
• Integrated Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) refinement for parameter policy.
• Evaluated RL-DE on the CEC2017 benchmark test set and flexible job shop scheduling problems.

Main Results:

• RL-DE demonstrated robustness in high-dimensional optimization scenarios.
• Achieved superior performance compared to classical adaptive DE variants on benchmark tests.
• Successfully applied to flexible job shop scheduling, validating generalization to discrete optimization.

Conclusions:

• RL-DE offers a learnable, unified architecture for adaptive parameter control in DE.
• Provides an efficient and intelligent solution for expensive black-box optimization.
• Shows significant potential for engineering scheduling and combinatorial optimization tasks.