[STAT/AP] Julia Olkhovskaya: Challenges and recent progress in reinforcement learning for large-scale environments.

20 januari 2025 15:45 t/m 16:45 - Locatie: Hall D@ta | Zet in mijn agenda

Reinforcement learning (RL) has gained considerable attention recently as a framework for making sequential decisions in uncertain environments. While most of the existing research focuses on settings with a finite number of state-action pairs (the tabular setting) or simple models, like linear state-action value functions, it remains challenging to develop an RL algorithm that can efficiently manage large state-action spaces with more complex value functions. Recent studies have looked into nonlinear function approximations, with kernel ridge regression offering a natural step between linear models and neural networks. I will discuss the range of model complexities - from tabular representations to linear and nonlinear function approximations - and highlight some ongoing results and open problems in achieving statistical efficiency in these areas.