BEGIN:VCALENDAR
METHOD:PUBLISH
PRODID:www-math-fau-de//Events//DE
VERSION:2.0
BEGIN:VEVENT
SUMMARY:Modelling choices in model-based Reinforcement Learning (G. Ko
 ntes\, Fraunhofer IIS)
UID:4915b0fb-3785-4089-9bc5-4721d959f842
DESCRIPTION:Modelling choices in model-based Reinforcement Learning Sp
 eaker: Dr. Georgios Kontes Affiliation: Self-Learning Systems Group\, 
 Precise Positioning and Analytics Department\, Fraunhofer Institute fo
 r Integrated Circuits IIS Zoom link: Meeting ID: 975 4039 5456 \, Pass
 code: 910962 Abstract: Reinforcement Learning (RL) is an area of Machi
 ne Learning concerning with agents that take sequential decisions with
 in an environment\, aiming at solving a given problem. In the classica
 l Reinforcement Learning paradigm the learning agent has no knowledge 
 of the dynamics of the environment or labeled examples of correct acti
 ons\, but instead must interact with the environment to design a polic
 y (controller) that maximises future cumulative reward. Among the seve
 ral different approaches within the RL ecosystem\, model-based RL is o
 ne of the most sample-efficient\, providing also some system safety gu
 arantees. Here\, a model of the system dynamics is learned using regre
 ssion techniques and future actions are selected based on online plann
 ing algorithms that utilise the learned model. A core question here is
  which type of model to use: simpler models might not be expressive en
 ough for complex problems or adequate for large datasets\, while highe
 r-capacity models tend to overfit to low-data regimes. Another importa
 nt aspect is how to consider model uncertainty throughout planning\, t
 hus improving data efficiency during learning and safe operation durin
 g deployment. In this presentation\, different variants from the RL li
 terature for the model and the online planner will be presented. Appli
 ed Analysis
DTSTART:20200909T103000Z
DTEND:20200909T113000Z
LOCATION:Online
DTSTAMP:20260730T180825Z
END:VEVENT
END:VCALENDAR