Fwd: Reminder - Thesis Proposal - May 18, 2022 - Ian Char - Quick Adaptation via Latent Variables in Model-Based Reinforcement Learning for Nuclear Fusion Control

[Jeff Schneider]

Hi Everyone,

Please come and see Ian's thesis proposal on learning and nuclear 
fusion!  It starts in 28 minutes.

Jeff.



-------- Forwarded Message --------
Subject: 	Reminder - Thesis Proposal - May 18, 2022 - Ian Char - Quick 
Adaptation via Latent Variables in Model-Based Reinforcement Learning 
for Nuclear Fusion Control
Date: 	Tue, 17 May 2022 15:24:58 -0400
From: 	Diane Stidle <stidle at andrew.cmu.edu>
Reply-To: 	stidle at andrew.cmu.edu
To: 	ml-seminar at cs.cmu.edu <ML-SEMINAR at CS.CMU.EDU>, 
riedmiller at google.com, ekolemen at pppl.gov



/*Thesis Proposal*/

Date: May 18, 2022
Time: 9:30 AM (EDT)
Place: GHC 8102 + Remote
Speaker: Ian Char

*Title: Quick Adaptation via Latent Variables in Model-Based 
Reinforcement Learning for Nuclear Fusion Control*
*
*
Abstract:
Developments in machine learning suggest that learning controllers may 
soon be a viable alternative over the arduous process of developing 
controllers for complicated systems. Control for nuclear fusion is a 
perfect opportunity to employ these algorithms as the dynamics are 
highly non-linear and the physical models that do exist are difficult to 
use for control. In this thesis proposal, I will start by highlighting 
our efforts to learn controllers for fusion with both Bayesian 
optimization and model-based reinforcement learning. As is often the 
case in real world control problems, the performance of these 
controllers are degraded from the fact that the system dynamics of the 
device change from run to run. This is especially the case in nuclear 
fusion as each run of the tokamak may behave differently, which means 
the controller must be able to adapt quickly during operation. To 
address this, I propose modelling the system dynamics using latent 
variables which are observed by the controller and can be quickly 
adjusted for each new run of the device. This quick adaptation is also 
important in cases where the true dynamics are unknown, and I will 
highlight how better epistemic uncertainty modelling can be used to 
improve model-based reinforcement learning. Lastly, I will describe how 
these methods will be leveraged to create powerful controllers for the 
nuclear fusion community.

*Thesis Committee:*
Jeff Schneider (Chair)
Zico Kolter
Ruslan Salakhutdinov
Martin Riedmiller (DeepMind)
Egemen Kolemen (Princeton)

Zoom Meeting link:
https://cmu.zoom.us/j/95486052837?pwd=amNJWlJFUWpYUm9vOGFYVTYvUzduZz09

Link to draft document:
https://drive.google.com/file/d/1oyVdMq6Il2_JXQsDS1J_AkcaIgFP3sEk/view?usp=sharing

-- 
Diane Stidle
Graduate Programs Manager
Machine Learning Department
Carnegie Mellon University
stidle at andrew.cmu.edu