I know that reinforcement learning has been used to solve the inverted pendulum problem.

Can supervised learning be used to solve the inverted pendulum problem?

For example, there could be an interface (e.g. a joystick) with the cart-pole system, which the human can use to balance the pole and, at the same time, collect a dataset for supervised learning. Has this been done before?

  • 1
    $\begingroup$ The inverted pendulum is easily controlled with little effort. Using some machine learning method is surely a waste of time. Is this a research question or one of curiosity? $\endgroup$ Nov 28, 2019 at 12:09
  • $\begingroup$ I agree that an inverted pendulum is easily controllable. But what if we are talking about a plant which is not easily describable i.e does not have a transfer function or is a nonlinear system. The Plant dynamics can change over time. In such a scenario, can we use supervised learning? It is a question out of curiosity and if valid, can be for research too. $\endgroup$ Nov 28, 2019 at 14:24
  • $\begingroup$ You really need to decide what you’re asking. Are you asking about supervised learning for a pendulum, or for a non describable system. Decide which you want to ask for, then modify your question to reflect your actual question. To stick with what is actually being asked, it’s possible sure, generally one can do anything they want. But you’d make a control program to first keep the pendulum inverted, since people probably couldn’t do it by hand with a joystick anyways, then supervise it afterwards and use said data to teach your Learning system...but since its already controlled...why bother? $\endgroup$ Nov 28, 2019 at 14:29
  • $\begingroup$ I wanted to give an example of an inverted pendulum. In general, humans, through experience, can adapt to changing conditions intuitively, through very minimal iterations compared to reinforcement learning. So if I, for example, wanted to achieve precise position control of a motor, in varying conditions, and teach a machine to learn how to perform in the changing conditions, could it also achieve precise position control of the motor in these varying conditions? $\endgroup$ Nov 28, 2019 at 14:45
  • $\begingroup$ "the human can use to balance the pole and, at the same time, collect a dataset for supervised learning" That sounds more like imitation learning to me. $\endgroup$ Dec 3, 2019 at 19:15


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.