An improved differential evolution algorithm based on reinforcement learning and its application.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Swarm Intelligence

Background:

• Differential Evolution (DE) is a powerful swarm intelligence method for high-dimensional problems.
• DE suffers from parameter sensitivity and premature convergence, limiting its practical application.
• Existing optimization methods require improvements in efficiency and adaptability.

Purpose of the Study:

• To propose an improved Differential Evolution algorithm, RLDE, leveraging reinforcement learning.
• To enhance the global optimization performance and address limitations of the standard DE algorithm.
• To validate the algorithm's effectiveness on benchmark functions and a real-world engineering problem.

Main Methods:

• Population initialization using Halton sequence for improved ergodicity.
• Dynamic parameter adjustment via a reinforcement learning policy gradient network for adaptive scaling factor and crossover probability.
• Differentiated mutation strategy based on population fitness classification.

Main Results:

• RLDE significantly improved global optimization performance across 26 standard test functions.
• The algorithm outperformed multiple heuristic optimization algorithms in 10, 30, and 50 dimensions.
• Successful application to the Unmanned Aerial Vehicle (UAV) task assignment problem demonstrated practical engineering value.

Conclusions:

• The proposed RLDE algorithm effectively addresses DE's parameter sensitivity and premature convergence.
• RLDE offers enhanced global optimization capabilities for complex, high-dimensional problems.
• The algorithm shows significant potential for real-world engineering applications, such as UAV task assignment.