Robust Dynamic Material Handling via Adaptive Constrained Evolutionary Reinforcement Learning.

Area of Science:

• Operations Research
• Artificial Intelligence
• Robotics

Background:

• Dynamic material handling (DMH) requires real-time task assignment to minimize makespan and tardiness.
• Reinforcement learning (RL) shows promise for DMH but faces challenges like sparse rewards and constraint satisfaction.
• Adaptability to dynamic events and efficient use of historical data are critical for robust DMH policies.

Purpose of the Study:

• To propose a novel adaptive constrained evolutionary RL (ACERL) approach for dynamic material handling.
• To address challenges in DMH including sparse rewards, constraint violation, and efficient policy training.
• To improve the adaptability and robustness of decision policies in dynamic environments.

Main Methods:

• ACERL utilizes a population of actors for diverse exploration and to handle sparse rewards and constraint violations.
• The approach adaptively selects beneficial training instances to enhance policy learning.
• Extensive experiments were conducted on multiple training and unseen test instances, including noised scenarios.

Main Results:

• ACERL significantly outperforms state-of-the-art algorithms in dynamic material handling tasks.
• Trained policies successfully schedule vehicles while fully satisfying all constraints.
• Experiments demonstrate ACERL's outstanding performance, robustness on noised instances, and overall effectiveness via cross-validation.

Conclusions:

• The proposed ACERL approach offers a powerful solution for dynamic material handling problems.
• ACERL effectively balances performance optimization with strict constraint satisfaction.
• The adaptive and evolutionary components of ACERL are crucial for its robust and efficient operation.