Deep reinforcement learning algorithm for solving material emergency dispatching problem.

Area of Science:

• Operations Research
• Artificial Intelligence
• Supply Chain Management

Background:

• Dynamic changes in node demand disrupt real-time scheduling in emergency material dispatch.
• Existing scheduling schemes struggle to adapt to these real-time fluctuations, leading to inefficiencies.

Purpose of the Study:

• To propose a novel dynamic attention model for real-time scheduling in material emergency dispatching.
• To improve the adaptability and efficiency of dispatching schemes when faced with dynamic node demand.

Main Methods:

• A dynamic attention model based on an improved gated recurrent unit (GRU) was developed.
• A dynamic codec framework was employed to track and update node demand information.
• Node embeddings were generated by a weighted combination of historical and current node data.

Main Results:

• The proposed model demonstrated significant improvements in solution quality compared to the elitism-based immigrants ant colony optimization algorithm.
• Improvements ranged from 27.89% to 28.12% across varying problem scales (10 to 100 nodes).
• The model effectively mitigated instability caused by fluctuating node demands.

Conclusions:

• The dynamic attention model offers a robust solution for real-time scheduling in dynamic environments.
• This approach minimizes material distribution costs by effectively handling unstable node demands.
• The improved GRU enhances the model's representational capacity for better decision-making.