H.B. Keller Colloquium

As machine learning systems progress from prediction to decision making, we are faced with a critical challenge: How can we develop algorithms that can effectively train models using data from various users or environments, especially when these models will be deployed in dynamic, unseen environments where the next ‘state' is influenced by the algorithm's decisions? This is particularly relevant as most environments are not static; they are subject to unpredictable, often non-stationary uncertainties, as well as endogenous feedback-induced distribution shifts. As an example, there are already many large-scale machine learning systems that interact with users in operation, even though there is a significant gap in our theoretical understanding of how user responses to inputs---e.g., queries, recommendations, incentives, etc.---can trigger endogenous and exogenous dynamics rendering the training data obsolete or even adversarial.

In this talk, I will discuss some recent theoretical results on decision-making algorithms that are being designed and deployed in environments where there are multiple reactive users, ranging from myopic to strategic, in addition to time-varying stochastic uncertainties. I will present performance and efficiency analysis of decision-makers ranging from oblivious to strategic, and present learning algorithms with theoretical guarantees in such settings.  Time permitting, towards the end of the talk, I will delve into open questions and highlight some significant gaps between theory and practice for using learning-enabled components in complex decision-systems.