Évènement

Reinforcement Learning (RL) provides a framework for sequential decision-making under uncertainty, grounded in Markov Decision Processes (MDPs) and dynamic programming principles. This talk begins by reviewing the core concepts of RL, including MDPs, the Bellman equation, and classical algorithms for finding optimal policies.

In the second part, we shift the focus to risk-sensitive objectives. While standard RL maximizes expected returns, many real-world applications must account for variability and worst-case outcomes. We therefore introduce risk measures (e.g., Value at Risk or Conditional Value at Risk, Entropic Risk Measure) and discuss how incorporating them alters the Bellman equation and challenges existing theoretical guarantees.