CAREER: Enhancing Human-AI Collaboration Toward Trustworthy Learning-Based Network Controllers

Machine learning-based approaches are increasingly replacing manually designed heuristics in network controllers such as those used for adaptive bit rate video streaming and for Internet congestion control. However, these black-box learned models introduce new challenges for network operators. Key concerns include selecting appropriate training data, testing under realistic and variable network conditions, ensuring debuggability, and maintaining safety during online deployment. This project aims to develop novel explainability tools and techniques to enable effective human-AI collaboration during the design and deployment of learning-based network controllers. This project plans to pursue three synergistic thrusts to realize trustworthy learning-based network controllers. First, this project focuses on developing a concept-based explainer that interprets model behavior using high-level, human-understandable concepts that are easy for an operator to understand. Second, it plans to create a hybrid explainability framework that integrates concept-based, low-level feature-based, and predictive future performance-based explanations to provide a comprehensive and multi-layered understanding of controller decisions. Third, the project will investigate how these explanations can enhance key operational tasks, including data curation for training, test coverage evaluation, debugging, and root cause analysis. Together, these thrusts aim to bridge the gap between black-box models and practical network management needs. The domain-specific explainability solutions developed in this project will offer high-level insights into learned models, enabling network operators to better understand, trust, and manage learning-based controllers in real-world networks. By improving interpretability, these tools aim to lower the barrier to deploying high-performance learning-based solutions in practical network environments. In addition, they are expected to enhance the robustness and reliability of these controllers, enabling safer adoption of learning-based solutions in critical network applications. This project also involves close collaboration with industry partners and the development of a web-based platform to promote broader engagement, dissemination, and adoption of the developed tools. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.