Experienced Gray Wolf Optimization Through Reinforcement Learning and Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Optimization Algorithms

Background:

• Gray Wolf Optimization (GWO) faces challenges in parameter tuning, particularly the exploration/exploitation rate, impacting performance.
• Global parameter settings in GWO may not suit individual agent needs or dynamic search space conditions.

Purpose of the Study:

• To propose an enhanced GWO variant (EGWO) that utilizes reinforcement learning and neural networks for adaptive parameter control.
• To overcome the limitations of fixed parameter settings in GWO by enabling individual agent adaptation of exploration rates.

Main Methods:

• Developed EGWO by integrating reinforcement learning principles with neural networks to manage the exploration/exploitation rate on an individual agent basis.
• Created an experience repository using neural networks to map agent states to actions, dynamically adjusting exploration rates based on agent experience and search space characteristics.
• Updated the experience repository collaboratively by all agents to continuously refine future actions.

Main Results:

• EGWO demonstrated superior performance compared to the original GWO across various datasets.
• The proposed EGWO outperformed other metaheuristics, including genetic algorithms and particle swarm optimization.
• Evaluated EGWO's effectiveness in feature selection and neural network weight optimization using multiple performance indicators.

Conclusions:

• EGWO offers a significant advancement over traditional GWO by enabling adaptive, agent-specific parameter control.
• The integration of reinforcement learning and neural networks provides a robust framework for enhancing metaheuristic optimization.
• EGWO shows strong potential for application in complex optimization problems like feature selection and neural network training.