Imitation-regularized Optimal Transport on Networks: Provable Robustness and Application to Logistics Planning

Koshi Oishi,Yota Hashizume,Tomohiko Jimbo,Hirotaka Kaji,Kenji Kashima

2024-02-28

Abstract:Network systems form the foundation of modern society, playing a critical role in various applications. However, these systems are at significant risk of being adversely affected by unforeseen circumstances, such as disasters. Considering this, there is a pressing need for research to enhance the robustness of network systems. Recently, in reinforcement learning, the relationship between acquiring robustness and regularizing entropy has been identified. Additionally, imitation learning is used within this framework to reflect experts' behavior. However, there are no comprehensive studies on the use of a similar imitation framework for optimal transport on networks. Therefore, in this study, imitation-regularized optimal transport (I-OT) on networks was investigated. It encodes prior knowledge on the network by imitating a given prior distribution. The I-OT solution demonstrated robustness in terms of the cost defined on the network. Moreover, we applied the I-OT to a logistics planning problem using real data. We also examined the imitation and apriori risk information scenarios to demonstrate the usefulness and implications of the proposed method.

Machine Learning,Systems and Control

What problem does this paper attempt to address?

### What problems does this paper attempt to solve?

This paper aims to solve the robustness problem of network systems (such as supply chains, logistics, power grids and communication networks) when facing unforeseen situations (such as disasters). Specifically, the paper introduces a new method - **Imitation - regularized Optimal Transport (I - OT)** to enhance the robustness of network systems and applies it to the logistics planning problem.

#### Main problem description:

1. **Robustness of network systems**:
- Modern society depends on various network systems, which are vulnerable to unforeseen events. Therefore, it is crucial to improve the robustness of these systems.
- The paper points out that although existing research has explored enhancing robustness through entropy regularization, it has not fully explored applying the imitation learning framework to the optimal transport problem on networks.

2. **Optimal transport and robustness**:
- Optimal Transport (OT) is a classical mathematical tool for finding the minimum - cost path from one distribution to another in a network.
- The paper proposes a new method - I - OT, which enhances the robustness of the transport plan by introducing an imitation - regularization term to encode prior knowledge.

3. **Challenges in practical applications**:
- In practical applications, especially in the case of large - scale networks and non - Markovian structures, computational complexity and time cost are important challenges.
- The paper not only proposes theoretical improvements but also explores how to apply the I - OT method in practical logistics planning.

#### Specific contributions:

1. **Provable robustness**:
- Through mathematical derivation, the paper determines the magnitude of cost fluctuations that can guarantee robustness.

2. **Utilizing prior knowledge**:
- A method is proposed to incorporate information when it is known that the costs of some edges fluctuate greatly.

3. **Fast solution method for large networks**:
- A fast solution method suitable for large - scale networks is explored, which improves computational efficiency.

4. **Handling non - Markovian cost structures**:
- A method for dealing with networks that do not satisfy the Markov property in the real world is proposed.

### Summary

By introducing Imitation - regularized Optimal Transport (I - OT), this paper solves the robustness problem of network systems when facing unforeseen situations and successfully applies it to the actual scenario of logistics planning. This method not only enhances the robustness of the transport plan but also makes significant progress in computational efficiency and practical applications.