I'm interested in predicting a temporal variable $X(t)$ using a recurrent neural network. The target $X(t)$ depends on both itself at previous times as well as another temporal function $W(t)$. I always know and can compute $W(t)$ for any given time. My goal with the RNN is to build a model for $X(t)$.

I'm familiar with the general idea of a RNN, but have never used one in practice. I'd like to get an answer to a basic question about RNN training before I spend more time understanding more about the underlying details. If a RNN approach is incompatible with what I'm trying to do I'd like to know it upfront.

My question is, can I train the RNN if my data set only contains $X(t)$ at certain multiples of the time discreteness I'd like an answer to? For example, I want the RNN to produce an output for $X(t)$ for every time step $\Delta t$. If my data set for training only contains $X$ values at ${X(0), X(2 \Delta t), X(5 \Delta t), ...}$ and so on, does that pose a fundamental problem in training? I have $W(t)$ for all times (e.g., $W(i*\Delta t)$ for all $i$), but simply do not have the corresponding $X(t)$ values for all times.

I don't know if this is absolutely not a problem for training a RNN, if this is something which is a problem but for which there are workarounds, or whether this is a fundamental problem that means that a RNN might not be used as stated above.

Could someone please shed some light on this before I spend too much time pursuing this approach?


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.