DMARS_WGO: a deep reinforcement-driven hybrid metaheuristic for intelligent adaptive optimization

Area of Science:

• Optimization algorithms
• Computational science
• Engineering applications

Background:

• Metaheuristics are crucial for complex, high-dimensional problems.
• Existing algorithms struggle with exploration-exploitation balance, leading to premature convergence.
• Need for improved adaptive decision-making in optimization.

Purpose of the Study:

• Introduce two novel reinforcement-based metaheuristics: AIRE_WGO and DMARS_WGO.
• Enhance adaptive decision-making and convergence properties of optimization algorithms.
• Improve performance on complex scientific and engineering problems.

Main Methods:

• Developed Adaptive Intelligent Reinforced Walrus-Gazelle Optimizer (AIRE_WGO) using Q-learning for adaptive parameter control and diversity-informed mutations.
• Introduced Dual-Mode Adaptive Reinforced Switching Walrus-Gazelle Optimizer (DMARS_WGO) with a dual-agent reinforcement framework (Q-learning and Deep Q-Network).
• Implemented cross-agent knowledge sharing for enhanced cooperative intelligence and stability in DMARS_WGO.

Main Results:

• DMARS_WGO outperformed nine state-of-the-art optimizers on CEC2017 and CEC2022 benchmark suites and engineering design problems.
• DMARS_WGO achieved first rank in 26/29 CEC2017 functions and 8/12 CEC2022 functions.
• Statistical tests confirmed DMARS_WGO's significant superiority and robust self-adaptive search dynamics.

Conclusions:

• AIRE_WGO and DMARS_WGO offer advanced solutions for complex optimization problems.
• DMARS_WGO exhibits exceptional performance and robustness due to its dual-agent reinforcement learning and adaptive switching capabilities.
• The proposed algorithms, particularly DMARS_WGO, are highly effective for real-world engineering optimization tasks.