A review of reinforcement learning based hyper-heuristics.

Area of Science:

• Optimization Algorithms
• Artificial Intelligence
• Computational Intelligence

Background:

• Reinforcement learning based hyper-heuristics (RL-HH) integrate hyper-heuristics (HH) and reinforcement learning (RL) for dynamic strategy adjustment.
• RL-HH demonstrates effectiveness in solving complex real-world optimization problems.
• A comprehensive review and framework for the emerging RL-HH field are currently lacking.

Purpose of the Study:

• To provide a comprehensive review of existing reinforcement learning based hyper-heuristics.
• To present a general framework for understanding and categorizing RL-HH algorithms.
• To identify current research gaps and outline future research directions in RL-HH.

Main Methods:

• Systematic literature review of existing reinforcement learning based hyper-heuristics.
• Development of a general framework for RL-HH.
• Categorization of RL-HH algorithms into value-based and policy-based approaches.
• Detailed summary and description of typical algorithms within each category.

Main Results:

• Identification and categorization of current reinforcement learning based hyper-heuristics.
• Presentation of a generalized framework for RL-HH.
• Detailed descriptions of value-based and policy-based RL-HH algorithms.
• Analysis of existing research shortcomings and future research avenues.

Conclusions:

• Reinforcement learning based hyper-heuristics represent a significant advancement in optimization.
• The proposed framework facilitates a structured understanding of the RL-HH landscape.
• Further research is needed to address identified shortcomings and explore new directions in RL-HH.