Computer science is not my major, hence my question is two folded :

  • how is defined the problem I have (if I can name it, it will be easier for me to look for references) ?
  • what kind algorithm would be more suited to solve it ?

Let's take a wheeled remote controlled turtle (the real turtle) and a numerical model of it (the model turtle).

We first play once with the real turtle, record the commands and the resulting trajectory.

The problem is to find the right commands for the model turtle in order to fit (as much as possible) with the trajectory obtained previously with the real turtle.

With a perfectly representative model, it won't be an issue at all: for similar commands, one would get similar trajectory. But the model turtle will never be perfect, and commands will need to be tweaked.

Extra constraints are :

  • model turtle can only be initialized (at t=0) and run forward (in time) to the end of the trajectory (no reverse play, snapshot/recall, etc.). Simulation can be run many times though.
  • model turtle is in fact a complex black box system, with no known equations of it and a bunch of internal states (mostly filters).

Thanks in advance for the leads

  • $\begingroup$ It seems that you need a control loop, although normally control loops are used to make the physical system behave like the model system rather than vice versa $\endgroup$ – Vanessa Jul 6 '12 at 17:33
  • $\begingroup$ With a PID feedback we couldn't obtain good results. Mostly because the link between controls and trajectory are difficult to assess (behavior is deterministic but unknown a priori). Hopefully, I can re-play many times the model (use genetic algorithm ? but how to converge fast ?). I could also use a neural network as an input for a PID, but I have little idea on the best inputs and outputs for the NN. $\endgroup$ – yota Jul 11 '12 at 10:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.