FUN With Pendulum #6, Reinforcement Learning for the Inverted Double Pendulum

preview_player
Показать описание
Hey guys,

this video illustrates the learning process of an artificial agent that learns to balance an inverted double pendulum. The problem of training a software agent that learns intelligent behaviour in a particular environment is addressed in a subfield of machine learning called reinforcement learning. The exact learning strategy I used is called neuroevolution, which is one of the simplest ways to face the exploration-exploitation dilemma.

If you are interested in the equations of motion I used and how exactly I derived them, you find it in the following pdf in section 3.3.2.

Acknowledgements:

Music: The Workers, a Swiss funk band with me on the keys.

Music: Diledadafish, contemporary Swiss folk music with the sense of dadaistic humour and the feeling of red garlic.

Video: Meret, my beloved girlfriend, who helped me with the video recording.
Рекомендации по теме
Комментарии
Автор

Now you need an arduino, and a way to measure the angle of the joint. This should definitely become a physical project. What was it written in?

dataphreak
Автор

Nice work! I came here because of a reddit post (maybe it was yours) and this is a very interesting and hard worked video

Kernel-
Автор

Loved this! I loved watching the lines on the graph as much as the double pendulum 🙂

carnright
Автор

"Pendulum's mating dance" :)
Amazing work and great illustration!
Keep making fun videos man

goksanisil
Автор

Excellent work.
Is there a blog post where the whole process is explained in more detail?
Thanks for the contribution.

jsbaez
Автор

Good job!
Could you maybe please somehow release the code so people like me trying to learn machine learning can try to understand it?

mp_
Автор

how are you parameterizing the inputs to your neural net? is it the angles of the two joints? or is it the torque on the joints from gravity?

ElTurbinado
Автор

what exactly is the difference between blue and red pendulums in the simulations?

gauravs
Автор

question: I am relatively new to RNN's. Why would you need to derive the eq's of motion? So you can code the simulator (double pendulum)? So you can set optimal number of outputs? thanks in advance.

dynamitesteel
Автор

What software did you use to make the animations?

PlainPlaneOfficial