4 ms·
This looks like a great resource! Differentiable programming is such a cool area. I'll never forget writing a differentiable PID controller a few years ago and
by ubj 3y ago
This looks like a great resource! Differentiable programming is such a cool area. I'll never forget writing a differentiable PID controller a few years ago and watching the PID gains get tuned automagically to stabilize the control system. It's powerful stuff if you use it in the right places.
- persnickety 3y agoHow did you do it? Is there a resource explaining the process online? If there isn't, and you're willing to create one, I offer reviewing.
- redadehy2 3y agoAt Collimator.ai we built a differentiable modeling and simulation platform for dynamical systems on JAX (think Simulink but in the cloud and with Python+JAX instead of Matlab) in order to enable more efficient optimization, autotuning, MPC, neural network control, etc. Here is a simple example of PID tuning with JAX and autodiff: https://py.collimator.ai/examples/pid_tuning/ https://py.collimator.ai/examples/pid_tuning/ (feedback is welcome!)
- ubj 3y agoI haven't ever written up a tutorial, but here's a technical report from someone else that gives an overview: https://lucris.lub.lu.se/ws/portalfiles/portal/61129581/autotuning.pdf https://lucris.lub.lu.se/ws/portalfiles/portal/61129581/auto...