Deep Reinforcement Learning for Fluid Dynamics and Control

Поделиться
HTML-код
  • Опубликовано: 16 ноя 2024

Комментарии •

  • @Eigensteve
    @Eigensteve  3 года назад +24

    Links to papers in video:
    @3:58 Machine learning for fluid mechanics
    Brunton, Noack, Koumoutsakos, Ann. Rev. Fluid Mech 52:477--508, 2020
    www.annualreviews.org/doi/pdf/10.1146/annurev-fluid-010719-060214
    @5:04 Efficient collective swimming by harnessing vortices through deep reinforcement learning
    Verma, Novati, Koumoutsakos, Proc. Nat. Acad. Sci. 115(23):5849--5854, 2018
    www.pnas.org/content/115/23/5849
    @6:57 Automating turbulence modelling by multi-agent reinforcement learning
    Novati, Lascombes de Laroussilhe, Koumoutsakos, Nat. Mach. Int. 3:87--96, 2021
    www.nature.com/articles/s42256-020-00272-0
    @8:47 A review of Deep Reinforcement Learning for fluid mechanics,
    Garnier, Viquerat, Rabault, Larcher, Kuhnle, Hachem, 2019
    arxiv.org/abs/1908.04127
    @9:57 Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control
    Rabault, Kuchta, Jensen, Reglade, Cerardi, J. Fluid Mech. 865, 2019
    doi.org/10.1017/jfm.2019.62
    @10:56 Reinforcement learning for bluff body active flow control in experiments and simulations
    Fan, Yang, Wang, Triantafyllou, Karniadakis, Proc. Nat. Acad. Sci. 117(42), 2020
    doi.org/10.1073/pnas.2004939117
    @11:50 Fluid directed rigid body control using deep reinforcement learning
    Ma, Tian, Pan, Ren, Manocha, SIGGRAPH 2018
    gamma.cs.unc.edu/DRL_FluidRigid/
    @13:26 Autonomous helicopter flight via Reinforcement Learning
    Ng, Kim, Jordan, Sastry, NeurIPS 2004
    papers.nips.cc/paper/2003/file/b427426b8acd2c2e53827970f2c2f526-Paper.pdf
    @13:26 An Application of Reinforcement Learning to Aerobatic Helicopter Flight
    Abbeel, Coates, Quigly, Ng, NeurIPS 2007
    proceedings.neurips.cc/paper/2006/file/98c39996bf1543e974747a2549b3107c-Paper.pdf
    @13:26 Autonomous helicopter aerobatics through apprenticeship learning
    Abeel, Coates, Ng, Int J Rob Res 2010
    journals.sagepub.com/doi/abs/10.1177/0278364910371999
    @14:02 Learning to fly like a bird
    Tedrake, Jackowski, Cory, Roberts, Hoburg, Int. Symp. Rob. Res. 2009
    groups.csail.mit.edu/robotics-center/public_papers/Tedrake09.pdf
    @14:58 Control of a Quadrotor with Reinforcement Learning
    Hwangbo, Sa, Siegwart, Hutter, IEEE Rob Aut 2(4) 2017
    arxiv.org/abs/1707.05110
    @15:22 Learning to soar in turbulent environments
    Reddy, Celani, Sejnowski, Vergassola Proc. Nat. Acad. Sci. 113(33, 2016
    www.pnas.org/content/113/33/E4877
    @16:31 Learning to Fly: Computational Controller Design for Hybrid UAVs with Reinforcement Learning
    Xu, Du, Foshey, Li, Zhu, Schulz, Matusik, SIGGRAPH 2019
    people.csail.mit.edu/jiex/papers/LearningToFly/index.html

    • @vinca43
      @vinca43 Год назад

      You mentioned that your wife is a neuroscientist and indicated that animal reward is internal. Can you by chance provide a reference article that would describe this in greater detail?

  • @alexdosreisdesouza5916
    @alexdosreisdesouza5916 3 года назад +10

    It is so pleasant to see someone doing their work so passionately! You are an outstanding professor, Dr Brunton!

  • @aleph285
    @aleph285 3 года назад +2

    We need more professors and lecturers like you Dr. Brunton. You made academic publications more interesting to study !

  • @yusefalimam130
    @yusefalimam130 3 месяца назад

    you're so incredibly apt at explaining such mind-blowing phenomenon while tying them into interesting and novel areas of study! Thank YOU!

  • @bhargavkartik7476
    @bhargavkartik7476 3 года назад +4

    This channel is pure gold.

  • @priyanshukumawat4142
    @priyanshukumawat4142 3 года назад +3

    the best teacher I havve ever seen on youtube . greetings and regards from INDIA !!!!

  • @MichaelEvans-yq7xj
    @MichaelEvans-yq7xj 3 года назад +2

    I'm thinking about Jim Lovell and Apollo 13. After the explosion, he had to relearn flying the craft in a new configuration. He said it would go left when he wanted to go right. But they did it. This kind of work could save lives when we suddenly find ourselves in a new place we didn't count on.

    • @Eigensteve
      @Eigensteve  3 года назад +3

      I love these kinds of examples, where humans are able to rapidly re-learn on the fly.

  • @tonnirvana
    @tonnirvana 3 года назад +2

    The legend of control engineers 🙏 thanx

  • @shahriyarghazanfari4620
    @shahriyarghazanfari4620 2 года назад

    I really enjoy the enthusiasm you show when delivering the topics! Perfect and outstanding...

  • @khuongnguyenduy2156
    @khuongnguyenduy2156 3 года назад +5

    Thank you very much for sharing you great papers and knowledge!

  • @chanochbaranes6002
    @chanochbaranes6002 3 года назад +28

    It makes me want to join Washington University

    • @harv609
      @harv609 3 года назад +7

      I'm enrolled in prof. Brunton's April offering at UW, "machine learning for fluids". Will let you know how it is

    • @kevalan1042
      @kevalan1042 3 года назад +2

      @@harv609 is this offered as an online course?

    • @Eigensteve
      @Eigensteve  3 года назад +2

      Awesome, apply to our program in mechanical engineering!

    • @harv609
      @harv609 3 года назад

      @@kevalan1042 most courses at UW are offered remotely because of covid, so yea. It's offered as an online course this term

    • @kevalan1042
      @kevalan1042 3 года назад +1

      @@Eigensteve are you planning to offer the course on a MOOC platform such as Coursera or edX?

  • @joshuatienda
    @joshuatienda 3 года назад +1

    Thank you professor, it's great to watch your videos!

  • @kbrajasekarababu3878
    @kbrajasekarababu3878 2 года назад

    Dear prof,
    I really admire it.. Your videos make to do research on AI (DRL) in CFD
    Thank you professor.

  • @yankeesfan7052
    @yankeesfan7052 3 года назад

    Using the thermals to climb reminds me of the finite horizon, energy optimal trajectory video you just posted.

  • @dara_samii
    @dara_samii 3 года назад +3

    Fantastic Video
    please put the links of videos and papers you mentioned!

    • @Eigensteve
      @Eigensteve  3 года назад +2

      Great point. Just pinned a comment with links to all papers referenced.

  • @Mutual_Information
    @Mutual_Information 3 года назад +4

    Great video! One thing I'm curious about - my understanding is that reinforcement learning is difficult in practice b/c it's hard to build a simulated environment which matches reality. But here, it seems that issue has been managed, since these drones are flown in the real circumstance. So, are our turbulence simulations really that good or are there other clever tricks here? In general, do you see the accuracy of the simulation as the primary limiting factor in reinforcement learning?

  • @jeremyclark9697
    @jeremyclark9697 3 года назад

    This makes me really want to get into UW's CS program. Fingers crossed!

  • @jonathanviquerat3588
    @jonathanviquerat3588 3 года назад

    Hey Steve, thanks for the great video and the paper highlight ;)

  • @samuelleung9930
    @samuelleung9930 3 года назад +1

    Thank you, fantastic video!

  • @parthdevbundela6744
    @parthdevbundela6744 3 года назад +1

    Thank you for making this video

  • @syammohan2103
    @syammohan2103 3 года назад

    Great video and content!
    Would be really nice if you share how you are creating this video with content overlayed on screen.

  • @lorenzopaiola2908
    @lorenzopaiola2908 3 года назад +3

    So in the case of biological systems the reward is actually just a function of the state ?

    • @Eigensteve
      @Eigensteve  3 года назад +2

      Well it is somehow internally motivated, since it is your own body that is dumping the dopamine.

  • @commonwombat-h6r
    @commonwombat-h6r 3 года назад +1

    a great video! Thank you

  • @billykotsos4642
    @billykotsos4642 3 года назад +1

    This is so cool !
    RL fort the win !

  • @mohamedsalemfdahi3443
    @mohamedsalemfdahi3443 3 года назад +1

    Fantastic job big thanks

  • @abderrahmanechamsi3060
    @abderrahmanechamsi3060 3 года назад +1

    Amazing video thank you

  • @TriThom50
    @TriThom50 Год назад

    So these simulations (like in the case of the fish) are essentially used as RL environments, correct? Are these used within the framework of gym or what kind of toolset is used?

  • @motbus3
    @motbus3 3 года назад +1

    here we go!

  • @tantzer6113
    @tantzer6113 3 года назад +1

    Do you think these AI control techniques will replace current robust controller design techniques such as mu-synthesis, etc.?

    • @tantzer6113
      @tantzer6113 3 года назад +1

      I guess the advantage of traditional methodologies is that they come with mathematically proven guarantees on stability and performance, though no doubt they have disadvantages, too.

    • @Eigensteve
      @Eigensteve  3 года назад +1

      I don't think so, but they serve different purposes. Eventually I see RL leveraging more traditional control techniques more effectively.

  • @marcfreydefont7520
    @marcfreydefont7520 3 года назад

    Dear Sir, thank you for your channel and your book. These are a real treasure trove of very useful tools for the ‘honest mind’ of the 21st century. I was wondering whether you could direct me to resources regarding genetic algorithms. I have a couple of finance related optimisation problems involving discrete variables (essentially inclusion or exclusion in a portfolio subject to certain constraints) and I believe the genetic algorithms would be the most suitable in these cases. I would be extremely grateful for any guidance/suggestion on this. Many thanks and best regards from sunny Singapore.

  • @mohammadabdollahzadeh268
    @mohammadabdollahzadeh268 2 года назад

    Dear professor please prepare an education about using Q learning PID contriller

  • @jadsoj
    @jadsoj 3 года назад +1

    Steve, have you ever thought about creating a MOOC?

    • @Eigensteve
      @Eigensteve  3 года назад +2

      I have, but I like keeping everything 100% open on RUclips. But check out databookuw.com for more information on syllabi, homework, etc.

  • @sarvagyagupta1744
    @sarvagyagupta1744 3 года назад +1

    At 2:47 you said that in biological systems, the reward comes in the form of dopamine. But how is that dopamine released in the body? Wouldn't it be according to the environment we are in (or when we tell our mind that we are in a positive environment)?. Let me know.

    • @Eigensteve
      @Eigensteve  3 года назад

      Yes, but it is delivered internally, based on our internal perception of the external state.

    • @sarvagyagupta1744
      @sarvagyagupta1744 3 года назад

      @@Eigensteve but isn’t that equivalent to our brain thinking that we’ve got some positive response in the environment?

  • @oznerriznick2474
    @oznerriznick2474 3 года назад

    Well done!

  • @KevinWoogs
    @KevinWoogs 3 года назад

    Question: Are the fluid equations solved using a deep learning, or is it done through more traditional solvers like finite element?

  • @lgl_137noname6
    @lgl_137noname6 3 года назад +2

    11:42
    yeah, those people are called pitchers ...

  • @saeedsaimonable
    @saeedsaimonable 3 года назад

    Could u talking about architecture robot interactive/creative and AI

  • @mariogalindoq
    @mariogalindoq 3 года назад +1

    Probably SpaceX uses RL to land their rockets in the ships. It would be very nice to know something about that, however it should be secret.

    • @Eigensteve
      @Eigensteve  3 года назад +2

      Interesting... secret RL!

    • @seremetvlad
      @seremetvlad 3 года назад +3

      Pretty sure they use classical control. RL doesn't make a lot of sense in that particular case

    • @petraniccolai430
      @petraniccolai430 2 года назад

      @@seremetvlad could you explain why you think it doesn't make sense? I'm curious to understand

  • @alexanderlesnov2768
    @alexanderlesnov2768 7 месяцев назад

    Perfectly 🎉

  • @chymoney1
    @chymoney1 3 года назад +1

    Nice

  • @XecutionStyle
    @XecutionStyle 3 года назад

    Don't worry, we'll find the mystery behind this 'one person' solved by tomorrow 10'am.

  • @phaZZi6461
    @phaZZi6461 3 года назад +1

    e-fish-ently, i got it

  • @zhongzhongclock
    @zhongzhongclock 2 года назад

    3 unavailable videos are hidden?
    Doctor Brunton, so these 3 videos are still not open to public? But I still appreciate your first 5 videos which helps me understand reinforcement learning much more deeply. Thanks!