EEE 591/592 Seminar:
From Satisficing to Cones: Rethinking Sequential Decision-Making in the Era of Reality-Centric Machine Learning
Prof. Cem Tekin
Bilkent University
Date/Time: Monday, May 12, 2025 – 17:30-18:30 TSI
Place: EE 517 (in-person)
Abstract: Machine learning (ML) has been reshaping our world, driving breakthroughs from personal chatbots to life-saving clinical decision tools. Yet, as ML scales to real-world applications, it faces formidable challenges: ensuring fast learning, guaranteeing robust performance under unpredictable shifts, and balancing conflicting objectives. Sequential decision-making (SDM) plays a pivotal role in modern ML, with applications ranging from hyperparameter optimization to data-driven longitudinal patient treatment plans.
The first part of the talk aims to remaster SDM in ML by tackling fast learning and robustness from the lens of satisficing, an alternative decision-making model that, rather than seeking the absolute best choice, focuses on achieving a desired target. This approach speeds up learning and offers improved robustness guarantees in dynamic, imperfect environments compared to its robust and distributionally robust counterparts, making it particularly effective in “ML in the Wild” applications.
The second part of the talk proposes new algorithms to return parameters that trade-off conflicting performance objectives such as prediction accuracy and latency for expensive to evaluate black-box functions. These methods utilize preferences over objectives encoded by a polyhedral ordering cone to guide efficient search over the parameter space and come with strong performance guarantees on sample efficiency. The proposed algorithms, which can be viewed as black-box solvers of the celebrated vector optimization problem, have use cases beyond ML.
Lastly, I will discuss how satisficing and cones can be used in the broader context of reality-centric ML beyond SDM.
Biography: Cem received the BS degree in EEE from METU in 2008, graduating as valedictorian. He obtained his PhD in Electrical Engineering: Systems from the University of Michigan, Ann Arbor in 2013. Between 2013 and 2015, he worked as a postdoctoral researcher in the Department of ECE at the University of California, Los Angeles. He is currently an Associate Professor at Bilkent University.
Cem’s research focuses on exploration and exploitation in sequential decision-making problems. He has developed novel models and learning algorithms for Markov decision processes, contextual and combinatorial bandit problems, sequential multi-objective optimization, vector optimization, and Bayesian optimization. These algorithms have been applied to solve challenging engineering problems, ranging from personalized medicine to online recommendations to resource adaptation in cognitive communications. He has served as the principal investigator of five TÜBİTAK-funded projects and has served as an area chair in AISTATS. He has received several prestigious awards, including 2019 BAGEP Award, 2019 METU Parlar Foundation Research Incentive Award, 2020 IEEE Turkey Research Incentive Award, 2023 Bilkent University Excellence in Teaching Award, 2023 TÜBA-GEBİP Award, and 2024 TÜBİTAK Incentive Award.
Cem has published over 35 journal articles in leading journals (mostly IEEE Transactions) and presented over 35 conference papers at internationally renowned conferences, including AISTATS, ICLR, and NeurIPS. His research group, which currently includes 2 PhD and 5 MS students, continues to work on cutting-edge problems related to robustness, distributional shifts, and optimization of conflicting objectives.