Prof. Alexander Mathis's Lab: Alberto Chiappa "Novel methods to tackle adaptive reinforcement learning"
Online by invitation only
The ability to adapt is a distinguishing feature of intelligent systems. In nature, animals and humans prove remarkable adaptation skills, both when changes occur in the external environment and in their body. The sensorimotor system is a particularly relevant example, as it can often seamlessly recover task performance when perturbed.
Reinforcement Learning (RL) is the ideal learning framework to experiment with sensorimotor systems, because it enables the training of autonomous agents with complex control skills. While the classical formulation of an RL problem considered the same environment for training and testing, in recent years different testing frameworks, which evaluate the performance of an autonomous agent in the presence of perturbations, have emerged, opening new challenging research directions. I will talk about two methods for adaptive reinforcement learning: reinterpreting past experiences and learning state-dependent controllers.