Events

Postdoctoral Applicants

Toward Dependable Reinforcement Learning: Hidden Structures and Safe Optimization

Speaker: Ilyas Fatkhullin (ETHZ) Ph.D. Candidate, Department of Computer Science ETH Zürich Friday, February 27, 2026 1:00PM - 2:00PM and via Webcast: https://yale.zoom.us/j/99164580071?pwd=k5a3sTsnFsmStnzWO2VF2KhqjgiYR0.1

Zoom Password: 123

Abstract: Modern reinforcement learning (RL) has enabled impressive progress in sequential decision-making, yet state-of-the-art performance often requires enormous engineering efforts and still lacks reliable, theory-backed explanations. This talk develops an optimization-centric perspective on RL aimed at reliability: algorithms that remain stable in high-dimensional, stochastic, and inherently non-convex settings. I will walk you through three important pillars: (1) non-convex landscapes and hidden structures, (2) robustness and data-efficiency, and (3) scalable distributed learning. Going deeper into the first pillar, I show how RL objectives that appear intractable can nevertheless exhibit exploitable geometry — settings where non-convexity hides an implicit convex backbone. Such hidden convex structures yield global convergence guarantees for simple stochastic optimization procedures closely connected to policy-gradient-style methods. Building on this foundation, I address a critical challenge of safety in RL — respecting intrinsically non-convex constraints. I outline a globally convergent method based on a shifted inexact proximal-point framework that can drive iterates toward feasibility and optimality without relying on smoothness or constrained qualification assumptions, and demonstrate its practical behavior in safe exploration.
I conclude by outlining how (2) robustness and (3) scalability connect these geometric insights to reliable learning systems deployed under uncertainty and resource constraints, and discuss future directions.

Speaker Bio: Ilyas Fatkhullin is a Ph.D. candidate in Computer Science at ETH Zurich and an ETH AI Center Fellow. His research lies at the intersection of optimization theory and reinforcement learning, developing principled methods for reliable learning in high-dimensional, non-convex, and stochastic settings, including hidden convexity, heavy-tailed noise, and distributed training. His work has appeared in venues such as NeurIPS, ICML, and AISTATS, and in journals including SIAM Journal on Optimization, SIAM Journal on Control and Optimization, and JMLR. Special highlights include an oral presentation at NeurIPS 2021 (main track), a spotlight presentation at ICML 2022 (main track), and oral presentations at the NeurIPS 2025 OPT and COML workshops.

Add To: Google Calendar | Outlook | iCal File

• Postdoctoral Applicants