From Predictions to Decisions in Dynamic Strategic Environments

Abstract:
Many real-world games take place in environments that change over time due to external and sometimes adversarial shocks. Adapting strategies to these changes can be difficult for players. However, they often have access to predictions of future states of nature, for example from machine learning models. This presentation studies how such predictions can be used in non-stationary strategic environments. First, we introduce prediction-aware learning, a game-theoretic framework for players who rely on forecasts in time-varying games. We present learning algorithms that use predictions and provide guarantees on equilibrium convergence and social welfare. Second, we study how players should train their prediction models. Simply minimizing a standard prediction error is not always the best approach, since accurate predictions do not necessarily lead to good actions. This motivates decision-focused learning, where models are trained to directly improve the quality of the resulting decisions. We introduce an online decision-focused learning framework, along with algorithms, theoretical guarantees, and experiments. We illustrate these ideas with an application to power markets.

Biography:
Aymeric Capitaine is a third-year PhD student at École Polytechnique/Inria Paris, supervised by Michael I. Jordan, Alain Durmus, and Étienne Boursier. His research focuses on the strategic foundations of multi-agent systems. He revisits classical microeconomic models through the lens of modern learning methods and studies the interaction between prediction and decision-making in time-varying environments.