Reliable and crisis-resistant inventory optimization
The impact of the COVID 19 pandemic has raised public awareness of the importance of reliable supply chains. Empty supermarket shelves and drug shortages have shown the public how fragile modern supply chains are. From the consumer's perspective, it is therefore crucial that supply chains can withstand unforeseen fluctuations in demand, problems with suppliers and transport disruptions.
The focus of our work is on inventory optimization. The task is to automatically determine the most suitable inventory levels within a supply chain. The goal is to satisfy customer demand while minimizing the costs incurred by excess inventory. This is a classic application of supply chain optimization, often overseen by a supply chain manager.
Previous calculation methods for inventory optimization use static optimization methods. These strategies are based on maintaining a minimum inventory level that must be maintained at all times. A reorder quantity is ordered when the inventory falls below the defined threshold. Static methods are commonly used in practice because they are easy to communicate to supply chain stakeholders. However, they fail when the supply chain changes drastically because they lack the necessary flexibility to adapt.
In contrast, the more flexible machine learning methods, especially reinforcement learning, tend to be very complex and difficult to interpret and communicate to supply chain managers. This hinders their use in real-world scenarios, as they cannot be trusted without interpretability. The advantage of these methods is that we can create a crisis scenario in a simulator and prepare a strategy for difficult situations.
In our work, we combine the advantages of both methods by proposing a flexible, interpretable strategy that is learned using reinforcement learning. Our overall approach is shown in the figure above. A virtual supply chain manager, modeled as an interpretable machine learning model, is optimized by reinforcement learning to generate order quantities for each stage of a multi-echelon supply chain in a simulation. The agent observes the current state of inventory at each stage and retrieves the last actions to determine the next. Once deployed, the interpretable model can be exported as a graph, allowing the supply chain manager to examine how each feature in the observation space affects the order quantity decision.
In summary, our work bridges the gap between classical and modern inventory methods, combining their advantages and achieving comparable performance to less transparent inventory optimization methods. In future work, we will investigate more complex supply chains and multiple products.
Sponsored by Federal Ministry of Economics and Climate Protection.
The work was accepted as a workshop paper at the AI4ABM workshop at ICLR 2023.
Link to the paper: http://arxiv.org/abs/2303.10382
Subscribe to the Merantix Momentum Newsletter now.
More articles
.png)
%20(1).png)
.png)
.png)
Reliable and crisis-resistant inventory optimization
The impact of the COVID 19 pandemic has raised public awareness of the importance of reliable supply chains. Empty supermarket shelves and drug shortages have shown the public how fragile modern supply chains are. From the consumer's perspective, it is therefore crucial that supply chains can withstand unforeseen fluctuations in demand, problems with suppliers and transport disruptions.
The focus of our work is on inventory optimization. The task is to automatically determine the most suitable inventory levels within a supply chain. The goal is to satisfy customer demand while minimizing the costs incurred by excess inventory. This is a classic application of supply chain optimization, often overseen by a supply chain manager.
Previous calculation methods for inventory optimization use static optimization methods. These strategies are based on maintaining a minimum inventory level that must be maintained at all times. A reorder quantity is ordered when the inventory falls below the defined threshold. Static methods are commonly used in practice because they are easy to communicate to supply chain stakeholders. However, they fail when the supply chain changes drastically because they lack the necessary flexibility to adapt.
In contrast, the more flexible machine learning methods, especially reinforcement learning, tend to be very complex and difficult to interpret and communicate to supply chain managers. This hinders their use in real-world scenarios, as they cannot be trusted without interpretability. The advantage of these methods is that we can create a crisis scenario in a simulator and prepare a strategy for difficult situations.
In our work, we combine the advantages of both methods by proposing a flexible, interpretable strategy that is learned using reinforcement learning. Our overall approach is shown in the figure above. A virtual supply chain manager, modeled as an interpretable machine learning model, is optimized by reinforcement learning to generate order quantities for each stage of a multi-echelon supply chain in a simulation. The agent observes the current state of inventory at each stage and retrieves the last actions to determine the next. Once deployed, the interpretable model can be exported as a graph, allowing the supply chain manager to examine how each feature in the observation space affects the order quantity decision.
In summary, our work bridges the gap between classical and modern inventory methods, combining their advantages and achieving comparable performance to less transparent inventory optimization methods. In future work, we will investigate more complex supply chains and multiple products.
Sponsored by Federal Ministry of Economics and Climate Protection.
The work was accepted as a workshop paper at the AI4ABM workshop at ICLR 2023.
Link to the paper: http://arxiv.org/abs/2303.10382
.png)
.png)
