(NN Toolbox) Special recurrent network training regime with overshooting

Hi,

I want to implement a special network architecture/training regime and I have no clue how to do that. It would be great if somebody could point me in the right direction.

UPDATE (originally I tried to simplify the problem but it looks like I have to tell the whole story to make it plausible):

I want to have a dynamic NN that looks like this:

x(t), a(t) --> s(t) --> y(t)
            ^-------|

Intuition: I'm trying to do system identification of an unknown system (e.g. a robot).

• x(t): measurement of some sensor (e.g. a camera image of the robot) at time t
• a(t): action the sytem (e.g. robot) executes at time t
• y(t): boolean signal that says if the system has reached it's goal at time t
• s(t): hidden state of the system (this is what I am actually interested in)

the input x and a are mapped to a layer s that is self recurrent and then mapped to the target y. This is of course just an Elman-Network. The problem is the following:

I want to provide the network with the x's and a's for a certain time frame (say times T-5 until T) and then do overshooting: predicting y for the following (say 10) timesteps (from T+1 to T+10) with only providing the a's (not the x's anymore).

Intuition: In my problem, computing y(t) from x(t) is a trivial problem, the hard problem is predicting how y behaves from the current hidden state s(t) and the future actions a. To predict that, there needs to be some information in s that allows such predictions.

In the actual application I am only interested in estimating the current state sof the system given the past inputs x and a: x(T-5:T), a(T-5:T) --> s(T)

I hope this is not too confusing... It would be great if somebody has an idea how to do that!

Rico

Answer 1

[Greg Heath]

What you ask makes no sense to me.

If you want a net eventually predicts without an exogeneous input, design an openloop narnet. Then convert it to a closed loop narnet and, if needed, use the openloop weights to initialize training of the closed loop configuration.

The only other choice is to design the closeloop narnet from scratch. However, this will take quite a bit longer.

What is your reason for wanting to begin with a non-narnet design???

Hope this helps.

Thank you for formally acceptingmy answer

Greg

[Rico]

Hi Greg,

Thanks for your quick response!

I thought I could simplify my problem and not have to explain the whole story but it looks like this is confusing, which I totally understand. I will edit the problem description to make things more clear.