Jonas Degrave, Federico Felici, Jonas Buchli
Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de las Casas, Craig Donner, Leslie Fritz, Cristian Galperti, Andrea Huber, James Keeling, Maria Tsimpoukelli, Jackie Kay, Antoine Merle, Jean-Marc Moret, Seb Noury, Federico Pesamosca, David Pfau, Olivier Sauter, Cristian Sommariva, Stefano Coda, Basil Duval, Ambrogio Fasoli, Pushmeet Kohli, Koray Kavukcuoglu, Demis Hassabis, Martin Riedmiller
2022 · Nature
Abstract Nuclear fusion using magnetic confinement, in particular in the tokamak configuration, is a promising path towards sustainable energy. A core challenge is to shape and maintain a high-temperature plasma within the tokamak vessel. This requires high-dimensional, high-frequency, closed-loop control using magnetic actuator coils, further complicated by the diverse requirements across a wide range of plasma configurations. In this work, we introduce a previously undescribed architecture for tokamak magnetic controller design that autonomously learns to command the full set of control coils. This architecture meets control objectives specified at a high level, at the same time satisfying physical and operational constraints. This approach has unprecedented flexibility and generality in problem specification and yields a notable reduction in design effort to produce new plasma configurations. We successfully produce and control a diverse set of plasma configurations on the Tokamak à Configuration Variable 1,2 , including elongated, conventional shapes, as well as advanced configurations, such as negative triangularity and ‘snowflake’ configurations. Our approach achieves accurate tracking of the location, current and shape for these configurations. We also demonstrate sustained ‘droplets’ on TCV, in which two separate plasmas are maintained simultaneously within the vessel. This represents a notable advance for tokamak feedback control, showing the potential of reinforcement learning to accelerate research in the fusion domain, and is one of the most challenging real-world systems to which reinforcement learning has been applied.
Metadata: Crossref. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
3 selected excerpts · 3 total mentions
> Every time i've actually tried something like this it has not outperformed statistical process control. There are clear counterexamples to your experience, most notably in maintaining plasma stability in tokamak reactors: https://www.nature.com/articles/s41586-021-04301-9
Could make breakthroughs on fusion, through materials research or something else, and it has shown some results on fusion confinement control systems: https://www.nature.com/articles/s41586-021-04301-9
I wouldn't be so quick. #1 is usually implemented by Deep Reinforcement Learning systems, RL is an AI paradigm descended from Control Theory and is used extensively elsewhere in Optimization and Operations Research. Here[0] for example is a Nvidia blog post detailing how they used it to obtain circuits with 25% less area at the same performance metrics. (I didn't search for this post except just to get the link, it's very typical of the kind of research I follow and I read it as soon as it was posted several weeks ago.) In another[1] example, Deepmind researchers used DRL to control a fusion reactor with results that exceed the current state of the art. I used games as…
examples just because they are the easiest to understand and most popular applications of RL, and also very general, but that would be like saying that CNNs are useless because recognizing cats is not a bussiness, recognizing cats is just 1 example of the vast array of things CNNs can do. For every optimization algorithm pioneered by RL research in order to play a useless $0-return game, you can bet that it's later implemented and used in other areas to save and generate millions or billions of dollars. Games are just the test playground for new research. I don't understand how #3 can't be credited with the billions of dollars the products and services based on it generate. According to [2], the global market of offering a speech-to-text API is valued at 1.3 billion dollars in 2019, if you assume that just 20% of that is actual customer value (to account for hype, errors in valuation, etc...), that's about 260 million dollar per year. Even an extremly conservative value estimate tells you the technology generates a billion dollar every 4 year (if it stays constant, the forecast projects the market will reach >3 billions by 2027). Keep in mind that this does not include any value generated from any other component or product, this is just the profits from the APIs alone, the generous 20% discount is to account for hype, analysis errors and bad data. But you know what ? the broader point is, these are just 3 examples that I pulled off the top of my head while I was making a 5 minute comment without consulting google. And considering the person I'm replying to never provided a single example or any kind of specific clarification till now, or anwered any of my questions or otherwise offered any kind of rebuttal, I think it's pretty fair to say they hold up nicely. [0] https://developer.nvidia.com/blog/designing-arithmetic-circu... [1] https://www.nature.com/articles/s41586-021-04301-9 [2] https://www.fortunebusinessinsights.com/speech-to-text-api-m...
More from the same topics.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Chiyuan Zhang, Samy Bengio, Moritz Hardt +2
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
Scott Aaronson
In honor of Alan Turing's hundredth birthday, I unwisely set out some thoughts about one of Turing's obsessions throughout his life, the question of physics and free will.
Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski +10
We trained a convolutional neural network (CNN) to map raw pixels from a single front-facing camera directly to steering commands.