How would I best deal with input to a neural network that is only available upon request, but with a limited frequency?

To be specific, I aim to train a driver for a Simulated Car Racing Championship (link to competition software manual) using deep reinforcement learning.
Available as input are a variety of sensors providing distances from the racing car to the edge of the track, of which some can be pointed jointly in a new direction every time, but only be used again after at least 50 tics.

The two parts of my question are:
How do I best deal with data that is available only very infrequently,
and is it feasible to have these inputs controlled by the neural network, and if so what challenges does that create and what would be a good approach?


I would try giving the NN access to a basic physics simulation that it can use to interpolate between data points. There is a formalization of this technique called Kalman filter.

  • $\begingroup$ Welcome to Cross Validated. Can you expand on your answer so that the OP can implement it? You can have a look at the tour to familiarize yourself with the philosophy of the site. $\endgroup$ Aug 22 '16 at 18:16
  • $\begingroup$ While this wouldn't be impossible, the data points are rather removed from the physical model and one of the main problems the NN is supposed to solve is building just that physical model from them. $\endgroup$
    – hyperfekt
    Aug 27 '16 at 13:06

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.