Re:Differentiable Control Problems
by freemint on Thursday August 27, 2020 @07:40AM
(#24292639)
I assume you are talking about the pendulum.
PID tuning is fundamentally restricted to holding set points and can't deal with non-linearites to reach those set points.
MPC if solved over a to small time horizon might fail to find a windup trajectory. But that isn't a restriction. However i would consider MPC overblown for this, as MPC is usually a series of optimal control problems.
By optimal control i assume you mean things approaches where you optimize a u() so it minimizes some integral. This approach does just that, only that u is not parametrized by polynomials/splines but by a neural network, taking state, time and other things into account.
MPC if solved over a to small time horizon might fail to find a windup trajectory. But that isn't a restriction. However i would consider MPC overblown for this, as MPC is usually a series of optimal control problems.
By optimal control i assume you mean things approaches where you optimize a u() so it minimizes some integral. This approach does just that, only that u is not parametrized by polynomials/splines but by a neural network, taking state, time and other things into account.