Lightweight Reinforcement Learning for Autonomous Adaptive Agents
Future intelligent agents need to autonomously act in rich dynamic environments, e.g., industrial plants, finance, healthcare, replacing or collaborating with humans in dangerous and tedious tasks. Reinforcement Learning (RL) is a well-established field of study to model how agents autonomously learn to act, which burst since the recent coupling with deep learning methods (deep RL), achieving groundbreaking results in previously unsolvable problems. Juxtaposed with the extraordinary results of deep RL is the realisation that its methods cannot be used out-of-the-box, especially for real-world problems, due to strong inefficiency issues and need of human supervision.
In this talk, the guest will be Carlo D'Eramo, an indipendent research gruoup leader of the LiteRL group at TU Darmastadt. He will describe his research focusing on the problem of how agents can efficiently acquire expert skills that explain the complexity of the real world. His research aims to answer this question by introducing lightweight (deep) RL methods to obtain autonomous and adaptive agents efficiently. He will explain his recent contributions, focusing in particular on the motivations behind each work and the significance of the results achieved. Through the description of his works, this talk has the further objective of providing a concise but effective overview of the current limitations of deep RL, and suggesting possible future research directions that are worth exploring.
For more information visit the website.