Decision-making based on data-driven models is becoming increasingly important as data collection improves and machine learning matures. Because data-driven models are often imprecise, recommended actions may appear good in simulation but fail when deployed. We develop decision-making methods that explicitly address model imprecision in reinforcement learning. In this talk, I will describe algorithms that leverage robust optimization to reduce the impact of model error on quality of decisions. I will address both model bias and variance and will discuss applications to problems in energy storage, agriculture, and others.
Marek Petrik is a Research Staff Member at the Solutions and Mathematical Sciences Department at IBM's T. J. Watson Research Center. He received his Ph.D. in Computer Science from the University of Massachusetts, Amherst. His research focuses on machine learning and optimization with a special interest in robust and risk-averse optimization, stochastic sequential optimization problems, and reinforcement learning. He has worked on applications that include agricultural and environmental monitoring, supply chain optimization, revenue management, and online recommendations.