Lecture 13: Adaptive Control
(Big slides) Download Big slides) (Small slides Download Small slides)
Learning for Dynamics and Control
For recent research performed in the field of Learning for dynamics and control (L4DC) you might find these conference links interesting, where many presentations are freely available
- 1st L4DC Conference 2019 Links to an external site. with videos Links to an external site.
- 2nd L4DC Conference 2020 Links to an external site.
- 3rd L4DC Conference 2021 Links to an external site.
- 4th L4DC Conference 2022 Links to an external site.
- 5th L4DC Conference 2023 Links to an external site.
Short-list of recommended talks:
-
Francesco Borrelli (UC Berkeley): "Learning to Predict and Control": https://www.youtube.com/watch?v=oxrpEGI7hhI Links to an external site.
-
Manfred Morari (University of Pennsylvania): "A Practitioner's Perspective: https://www.youtube.com/watch?v=tnBTn4oU2UU Links to an external site.
---------------------------------------------------------------------------------------------------------------------
Deep RL control of Tokamak Fusion Reactor (from Feb 2022):
- Video (you can skip the first 5min): https://youtu.be/cDQHcpv5l-8?t=315
Links to an external site.
- Paper with more details: Delgrave et al(Nature 2022) Download Delgrave et al(Nature 2022)
Personal reflection[BoB]: A main achievement is to have developed a so good model of the physical system that it can be used to get RL to work, and to then be trusted to run online.Note that it was built "on top of" a realtime control system doing most of the heavy work, so it was not learnt from scratch. The paper illustrates the power of combining new learning techniques with traditional modeling+control methods.
---------------------------------------------------------------------------------------------------------------------
RL in power systems.
---------------------------------------------------------------------------------------------------------------------
Solving the protein folding problem
- In some depth presentation of Deepmind's breakthrough with Alphafold2: https://www.youtube.com/watch?v=B9PL__gVxLI
Links to an external site.
- "Two-minute paper" (18 min long!) presentation of Alphafold https://www.youtube.com/watch?v=FYVf0bRgO5Q
Links to an external site.
Maybe unclear if it fits into "learning for control" yet, but it might in the future, when people control how to change and build these "molecules of life". It again illustrates the fact that the progress is possible by combining traditional domain expert knowledge, in feature representation and score function definition etc, with novel methods for learning.
---------------------------------------------------------------------------------------------------------------------
DeepMind's AlphaStar (2019)
- Video (8min) describing how Deepmind constructed a grandmaster level Starcraft II playing agent using RL: https://www.youtube.com/watch?v=jtlrWblOyP4
Links to an external site.
- More background and some game play can be found in these videos: https://www.youtube.com/watch?v=UuhECwm31dM
Links to an external site.
(5min) and https://www.youtube.com/watch?v=cUTMhmVh1qs Links to an external site.
(2+ hours)
- Details can be found in this blog post with paper Links to an external site.
---------------------------------------------------------------------------------------------------------------------
Learning to play Poker
- Video (5min) with https://www.youtube.com/watch?v=u90TbxK7VEA
Links to an external site.
Interesting because it is involves game theory with imperfect information
---------------------------------------------------------------------------------------------------------------------
Iterative Learning Control for Beam Control at ESS
- This presentation needs to wait until next year. Will be based on some work done here in Lund, where also people from our department have been involved.
--------- Not applications, but the big picture : -------------------------
NY Times February 2023: Bing's new chatbot left me deeply unsettled.
NY Times January 2023: ChatGPT Hijacks Democracy Links to an external site.
Nature February 2023: How will AI change mathematics? Links to an external site.
On the upcoming EU legislation for AI Download On the upcoming EU legislation for AI
--------- Not applications, but also worth a watch : -------------------------
Video about control of POMDPs: https://www.youtube.com/watch?v=yrC2O7T7aKU
Links to an external site.
Video about Model Free RL and TD Learning: https://www.youtube.com/watch?v=0iqz4tcKN58
Links to an external site.
And while you have a moment over, why not take this course in POMDPs using Julia and Pluto.jl Links to an external site. :-)