Given a sequence of numbers t(0) thru t(n), predict t(n+1)

[gcox]

I'm trying to take a sequence of dates (represented as numbers) and predict the next date, which is unknown at the time of prediction.

I expected to do this:

net.train([
  [date0, date1, ... dateN]
]);
net.run([date0, date1, ... dateN]); // -> Something near dateN + 1

Sadly, that throws an exception (TypeError: Cannot read property 'runInput' of undefined at LSTMTimeStep.run...)

Of course, all the examples I see here look like this:

net.train([
  [1, 2, 3]
]);
net.run([1, 2]); // -> 3

I don't understand why the last line isn't net.run([1, 2, 3]); // -> Something close to 4. I assume I'm missing the point of the TimeStep networks, but they are documented as being able to predict future values...which doesn't seem to be demonstrated by the examples IMO.

Actual question
Is what I'm trying to do possible with this library?

[robertleeplummerjr]

Can you throw your example into a jsfiddle for quick reproduction?

[gcox]

Sure, here you go: jsfiddle

[robertleeplummerjr]

You just need to add the length of the entire input + output. Example net.run([1, 2, 3], 4); // -> 3.435204029083252

[robertleeplummerjr]

This isn't well documented since the recurrent stuff was in beta, and is now nearly all rewritten. In either case, this should probably be added to documentation for future humans.

[gcox]

Ah, I see. I definitely misinterpreted that 2nd argument. Postponing my lobotomy for now.

Thanks for the quick attention!

[robertleeplummerjr]

Ty for the interest! Wait till we get GPU architecture released!