Achieving Robustness for Learned Systems Policies

Title of the Talk: Achieving Robustness for Learned Systems Policies
Speakers: Divyanshu Saxena
Host Faculty: Dr. Praveen Tamanna
Date: Dec 16, 2025

Abstract: Machine-learned policies for systems tasks—such as resource allocation, cluster scheduling, congestion control, and I/O prefetching—have shown substantial gains over traditional heuristics. Yet their adoption remains limited due to concerns about unsafe decisions and poor generalizability. In this talk, I will present case studies that apply techniques ranging from lightweight runtime monitors to analytical models and full-fledged verifiers to enhance the robustness of learned systems policies. Together, these case studies illustrate a broader design principle: learned policies can be made reliable when equipped with formal structure, domain knowledge, or runtime safeguards.

For the first case study, I will present evidence that learned microservice controllers fail in practice and that this brittleness originates from a lack of guidance on environmental conditions. I will then introduce `Galileo’, a framework that embeds queueing-theoretic reasoning into a Performance Reasoning Model to support what-if analysis of end-to-end performance under varying environments. I will illustrate how this analysis can be used both at train time, via reward shaping, and at inference time, via shielding. Galileo-trained autoscaling and admission-control policies reduce SLO violations by up to 99.4% on real-world benchmarks.

For the second case study, I will demonstrate that existing learned congestion controllers also frequently violate key safety and performance properties. I will present `Canopy’, a framework that provides provable property satisfaction for learned congestion controllers by leveraging traditional verifiers to supply corrective learning signals. Experimental results show that Canopy achieves up to 1.4× higher worst-case property satisfaction compared to the baseline Orca.

Finally, I will describe the challenges of integrating learning into the OS-kernel policies, and outline a vision for achieving robust operations via structured runtime `guardrails’ – runtime monitors that track properties of interest and take corrective actions.

This talk is based on works accepted to appear at NSDI’26, EuroSys’26 and HotOS’25.

Speaker Bio: Divyanshu Saxena is a Ph.D. candidate in the Department of Computer Science at UT Austin, advised by Prof. Aditya Akella. His research interests are broadly in the domain of Networked Systems, primarily focusing on microservice deployments and machine learning for systems. Before starting graduate school, he completed his B.Tech. in Computer Science and Engineering from IIT Delhi in 2020.