Study Focusing on Artificial Intelligence Applicability for ITS Identifies 12 Challenges and Offers Lessons and Insights to Help Mitigate These Challenges.
Nationwide, United States
Artificial Intelligence (AI) for Intelligence Transportation Systems (ITS) Challenges and Potential Solutions, Insights, and Lessons Learned
Summary Information
Artificial Intelligence (AI), including Machine Learning (ML), has the potential to create opportunities to bring transportation systems to more desirable levels in terms of safety, equity, reliability, accessibility, security, efficiency, and resilience. While recognizing the transformative potential of AI and ML, it's also crucial to be mindful of the challenges they present, such as data related issues, supporting technology, bias, security, privacy, ethics and equity, generalization, model drift, explainability, liability, talent/workforce availability, and stakeholder perception. In light of these observations, this study, led by USDOT's Intelligent Transportation Systems Joint Program Office (ITS JPO), focused on the implications of challenges surrounding AI implementation for ITS, offering insights that agencies could take into account to help mitigate these challenges.
- Share data with sufficient metadata and documentation. AI algorithms typically process vast volumes of data to generate inferences or predictions. Consequently, it's essential to provide practitioners with detailed documentation about the shared datasets, including their origins, nuances, and metadata to prevent any misinterpretation in the variables.
- Maximize data reuse to minimize redundancy and duplicate effort in similar use cases. However, the relevance of the use case should be carefully considered, as availability does not mean that they will be useful for a specific task.
- Explore synthetic, imputed, and human/AI collaborative approaches for data creation, and to perform thorough validation. This is because data manipulation and feature extraction using AI can often broaden the range of generalizable data points and enable the system to discern complex feature interrelationships.
- Ensure that training data is representative, manage edge cases, and limit overfitting for enhanced generalization. A ML model cannot extrapolate beyond the bounds of its training data labels, therefore, understanding the operational context of a ML model is crucial for ensuring the training data includes all possible patterns and outcomes. Performing model stress testing and developing anomaly detection can assist in identifying edge cases.
- Use model testbeds and combine ML techniques or models if needed. Utilizing model testbeds is recommended for testing AI system reliability without real-world consequences. Additionally, multiple learning techniques can be useful, especially when dealing with non-recurrent conditions.
- Develop a consistent plan to monitor system improvements and assess potential model drift. Regular testing of AI applications in ITS helps in detecting model drift and signals when retraining of the AI system is necessary should data or model drift surpass the established acceptable range.
- Pave the way with pilot deployments. Launching pilot deployments would serve the important purpose of uncovering potential barriers and demonstrating benefits. While doing so, it would also be helpful to leverage any existing ITS infrastructure where applicable to minimize costs and build sustainable AI solutions to reduce negative environment impacts.
- Apply cloud, edge and clustered computing to overcome computational, bandwidth and latency problems. Legacy systems used by many state and local agencies may have limited data storage and computational power. Using cloud, edge, and/or clustered computing could increase computational speed and augment processing power.
- Maintain a human-in-the-loop to support ongoing oversight of AI/ML applications in ITS. Humans can help in identifying and mitigating potential issues or nuances that the machine may not catch as well as those that may require making tradeoffs in how they are addressed.
