SUPAID: A Rule mining based method for automatic rollout decision aid for supervisors in fleet management systems

Sahil Manchanda, Arun Rajkumar, Simarjot Kaur, Narayanan UnnyarXiv 2020 · arXiv

Introduction

In public transit and logistics, the daily "rollout" decision—determining which vehicles are fit for service—is a high-stakes task. Supervisors must evaluate fleets where many vehicles possess minor pending defects. An incorrect decision can lead to mid-trip breakdowns, safety hazards, and increased maintenance costs. The SUPAID system (Supervisor Decision Aid) addresses this by using rule mining to identify subtle patterns of defects that human experts might overlook.

The Challenge of Human Expertise

Traditional machine learning models often aim to replicate human expert decisions. However, in fleet management, human supervisors are not infallible. While they are generally accurate when deciding to bench a vehicle (high precision for "Down" states), they often exhibit over-optimism when approving vehicles with multiple "minor" defects. This "one-way efficiency" means that training a model to simply mimic a supervisor also replicates their blind spots. SUPAID is designed to audit these decisions rather than just replicate them, specifically targeting the hidden risks in approved rollouts.

The SUPAID Methodology

The core of SUPAID is built on Frequent Item-set Mining using the FP-growth algorithm. The system processes historical vehicle data to find combinations of defects that correlate with future failures.

Data Representation and Labeling

The algorithm operates on a list of transactions sorted by vehicle ID and timestamp. Each transaction records the "Defect State"—the set of unresolved issues—and the vehicle's operational status.

Table 2: Example of vehicle transaction data showing defect states and status.

To identify risky patterns, SUPAID labels transitions between consecutive days. A transition is labeled "Remains available" ($\alpha$) if the vehicle stays operational, and "Becomes down" ($\beta$) if it fails after being approved for service.

Table 3: Example of labeled transitions used to train the model.

Scoring and Ranking

SUPAID calculates an Item Set Ratio (ISR) for various defect combinations. The ISR measures how much more likely a specific set of defects is to appear in the "Becomes down" group versus the "Remains available" group.

  1. Categorization: Historical data is split into sets $\alpha$ and $\beta$.
  2. Mining: FP-growth identifies frequent subsets of defects in both categories.
  3. Scoring: For a new rollout decision, the system evaluates all subsets of a vehicle's current defects and assigns a risk score based on the highest ISR found.
  4. Ranking: Vehicles are ranked by these scores, providing supervisors with a prioritized list of the safest vehicles to deploy.

Results and Impact

The system was validated using a year-long dataset from a major U.S. public transit agency, encompassing 249 vehicles and 945 unique defect types. The results demonstrated that SUPAID's data-driven approach significantly outperforms manual human decision-making.

  • Cost Efficiency: SUPAID achieved a 3.38% reduction in total maintenance costs compared to the supervisor's actual decisions.
  • Daily Savings: The average daily rollout cost dropped from $2,031.51** (human) to **$1,962.78 (SUPAID).
  • Interpretability: Unlike "black box" neural networks, SUPAID provides clear justifications for its rankings (e.g., identifying specific defect combinations with high failure rates), which is essential for gaining the trust of human supervisors.

Conclusion

SUPAID demonstrates that rule mining can effectively augment human expertise by identifying complex risk patterns that escape intuition. By focusing on the "one-way efficiency" of experts, the system provides a practical tool for reducing operational costs and improving fleet reliability. Future work will explore strategies to improve human adherence to algorithmic recommendations and multi-agent reinforcement learning for broader logistics optimization.

Summary updated

Abstract

The decision to rollout a vehicle is critical to fleet management companies as wrong decisions can lead to additional cost of maintenance and failures during journey. With the availability of large amount of data and advancement of machine learning techniques, the rollout decisions of a supervisor can be effectively automated and the mistakes in decisions made by the supervisor learnt. In this paper, we propose a novel learning algorithm SUPAID which under a natural 'one-way efficiency' assumption on the supervisor, uses a rule mining approach to rank the vehicles based on their roll-out feasibility thus helping prevent the supervisor from makingerroneous decisions. Our experimental results on real data from a public transit agency from a city in U.S show that the proposed method SUPAID can result in significant cost savings.
BibTeX
@misc{manchanda2020supaid,
  author = {Sahil Manchanda and Arun Rajkumar and Simarjot Kaur and Narayanan Unny},
  title = {{SUPAID: A Rule mining based method for automatic rollout decision aid for supervisors in fleet management systems}},
  year = {2020},
  eprint = {2001.03386},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.03386},
}
Figure