Most standard (reinforcement) learning settings focus on optimizing strategies for the expected return. Yet in risk-sensitive scenarios, moments beyond the mean can matter just as much—or even more. In some cases, the goal may be to avoid extreme outcomes (risk aversion), while in others, it may be to deliberately seek them out (playing for the lucky ticket).

In this talk, I will present two perspectives on learning beyond expectations. In the first part, I will show how extreme values provide a natural model for budget-constrained search problems, with applications to hyperparameter optimization. In the second part, I will discuss how the problem of planning in reinforcement learning becomes significantly more challenging once we move beyond simple expectations.

The talk will be based on two recent papers:

• Efficient Risk-sensitive Planning via Entropic Risk Measures

  Alexandre Marthe, Samuel Bounan, Aurélien Garivier, Claire Vernade (under review)

• Put CASH on Bandits: A Max K-Armed Problem for Automated Machine Learning

  Amir Rezaei Balef, Claire Vernade, Katharina Eggensperger (to appear at NeurIPS 2025)