A Self-Attention-Based Deep Reinforcement Learning Approach for AGV Dispatching Systems.

Area of Science:

• Operations Research
• Artificial Intelligence
• Robotics

Background:

• The automated guided vehicle (AGV) dispatching problem involves creating rules for assigning transportation tasks to vehicles.
• Existing methods may struggle in highly dynamic operational environments.

Purpose of the Study:

• To propose a novel deep reinforcement learning (DRL) approach with a self-attention mechanism for dynamic AGV task dispatching.
• To enhance the efficiency and adaptability of AGV dispatching systems in complex industrial settings.

Main Methods:

• Modeling the AGV dispatching system as a Markov decision process (MDP) with vehicle-initiated rules.
• Incorporating a self-attention mechanism to weigh information importance in dynamic environments.
• Utilizing invalid action masking and a multimodal structure for improved decision-making.

Main Results:

• The proposed DRL method demonstrates effectiveness in comparative experiments.
• Learned policies adapt to different system properties and effectively smooth traffic congestion.
• Under specific conditions, the policy converges to a shortest-queue-length heuristic, indicating adaptiveness.

Conclusions:

• The deep reinforcement learning approach with self-attention offers a powerful solution for dynamic AGV dispatching.
• The method shows significant adaptiveness and potential for optimizing logistics and traffic flow.
• This research contributes to intelligent automation in material handling systems.