Use Flux.Recur to run nonlinear simulation

[MrstupidJ]

RobustNeuralNetworks.jl/examples/src/lbdn_rl.jl

Line 48 in b7b1bee

md = Flux.Recur(forward, x0)

@raywang1987 I copied this piece of code (for running nonlinear system simulation) from our previous example (or maybe we did not use this, as the old example was linear), but it seems not working now.

[nic-barbara]

@MrstupidJ can you bit more specific on what the error is? Eg: provide a minimal example?

[MrstupidJ]

The Flux document on Flux.Recur was a bit confusing to control people. A simple example provided by the document was using an accumulation function as the cell:
accum(h, x) = (h + x, x)
rnn = Flux.Recur(accum, 0)
In Flux doc, h is the hidden states, x is the input to the cell.

To use Flux.Recur in control context, one can define a (discrete-time) nonlinear system dynamic as:
x, y = f(x, u)
(Note: the order of returned value matters, Flux will take the first as hidden states and second as output)
Then it can be initialized as a Flux model with initial state x0
model = Flux.Recur(f, x0)
Simulate one time step given input u(t):
y = model(u(t))
The hidden state (sate of the dynamic system) can be accessed in model.state.