Conversation

Whoa
Quote Tweet
Today in @nature, with @EPFL, the first deep reinforcement learning system that can keep nuclear fusion plasma stable inside its tokamaks, opening new avenues to advance nuclear fusion research. Paper: dpmd.ai/fusion-paper
Show this thread
Embedded video
0:02
332.1K views
2
44
This bit caught my eye… behind the jargon, the conventional approach is a *very* primitive strategy. SISO PID+gain scheduling is basically 1950s era controls. There would be no stability guarantees which removes one on the big reasons to use control theoretic over AI approaches.
Image
1
12
I always liked the idea of throwing AI at control problems, even though conservative friends would whine about lack of stability and convergence guarantees. But most controls as practiced in industry doesn’t use the advanced techniques that offer such guarantees anyway.
2
13
What’s more, most formally verifiable control architectures assume models that are not very good outside of very precisely modeled domains like aerospace. And tools to accommodate “model uncertainty” as it’s called are very limited (things like H-infinity never got very far)
1
3
Replying to
Yes, this is likely true. I also find it interesting how Deep Mind is very much working on real-world problems -- the folding of proteins to nuclear fusion. Compare that to OpenAI, seemingly only working on NLP. I wonder what Deepmind has in the pipeline?
1