An Off-Policy Reinforcement Learning-Based Adaptive Optimization Method for Dynamic Resource Allocation Problem.

Area of Science:

• Operations Research
• Artificial Intelligence
• Manufacturing Systems Engineering

Background:

• Dynamic resource allocation is critical in manufacturing but challenging due to multiple, often conflicting, objectives.
• Existing methods struggle with the complexity and adaptability required for real-time industrial environments.

Purpose of the Study:

• To propose an adaptive optimization method for multi-objective dynamic resource allocation problems (RAP) in manufacturing.
• To introduce a novel reinforcement learning approach, DSAC-ERCE, for enhanced performance in complex industrial settings.

Main Methods:

• Developed a novel deep state-actor-critic with entropy regularization and conditional entropy (DSAC-ERCE) reinforcement learning method.
• Implemented a state-encoding network with an information entropy attention mechanism for improved state representation.
• Introduced a new reward function to avoid local optima and a boundary method for high-quality actions.

Main Results:

• DSAC-ERCE demonstrated superior performance compared to state-of-the-art reinforcement learning methods in experiments.
• The method successfully adapted weights for multiple objectives and generated noninferior solutions dynamically.
• Validated generalization capabilities across various objective types (linear, quadratic, etc.) and problem structures.

Conclusions:

• The proposed DSAC-ERCE method offers an effective and adaptive solution for complex, multi-objective dynamic resource allocation in manufacturing.
• The approach shows significant potential for improving decision-making and operational efficiency in industrial settings.