Reinforcement learning for solution updating in Artificial Bee Colony.

Area of Science:

• Artificial Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• The standard Artificial Bee Colony (ABC) algorithm's one-dimension update process can degrade performance in high-dimensional optimization problems.
• This limitation affects solution quality and convergence speed, necessitating algorithmic improvements.

Purpose of the Study:

• To introduce a new algorithm, R-ABC, that integrates reinforcement learning into the ABC algorithm for enhanced solution updating.
• To address the performance drop observed in high-dimensional spaces by modifying the solution update mechanism.

Main Methods:

• The proposed R-ABC algorithm employs a reinforcement learning strategy within the onlooker bee phase.
• Positive or negative reinforcement is applied to solution dimensions based on fitness improvements from the employed bee phase.
• The update value for a dimension increases with frequent fitness improvements.

Main Results:

• R-ABC significantly outperformed other algorithms on basic numerical benchmark functions across various dimensions (100-900).
• Performance gains for R-ABC increased with higher dimensions on CEC2005 shifted functions.
• R-ABC demonstrated comparable performance to state-of-the-art ABC variants on CEC2014 hybrid functions.

Conclusions:

• The R-ABC algorithm effectively improves upon the standard ABC algorithm, particularly in high-dimensional optimization tasks.
• Reinforcement learning integration offers a promising direction for enhancing swarm intelligence algorithms.
• R-ABC presents a robust alternative for complex optimization problems where dimensionality is a challenge.