Fitted states for ESN?

[sdwfrost]

While there's an ESNpredict function, and it's possible to add the exact input state onto the state of ESN, would it be possible to add an `ESNfitted' function that returns the (predicted) state for the input/training data?

[MartinuzziFrancesco]

This would make for a very interesting addition! Instead of a new function I could add a return_states boolean to ESNpredict that the user would toggle if they want the actual prediction or the trained states. Even an option of having both could be useful. From an user's perspective do you think this could flow better?

[sdwfrost]

Arguments for a separate function is that for getting the trained states, all one needs is a call like ESNfitted(esn, W_out::AbstractArray{Float64}) - one doesn't need the prediction length, plus there's less chance of confusion in what the output is.

[MartinuzziFrancesco]

I am sorry, actually misread the initial post. I thought that you were looking for the fitted states for the prediction data. The fitted states for the input/training data can be accessed in the ESN struct: using the example in the readme one can fit the esn like so

esn = ESN(approx_res_size,
    train,
    degree,
    radius,
    activation = activation, #default = tanh
    alpha = alpha, #default = 1.0
    sigma = sigma, #default = 0.1
    nla_type = nla_type, #default = NLADefault()
    extended_states = extended_states #default = false
    )

and after to access the states you can call the states field of the ESN struct

esn.states
# 300×5000 Matrix{Float64}

[sdwfrost]

Hi, I think we're talking at crossed purposes. What I'd like is the readout for the input data (on the same scale as the observations), not just the internal states of the ESN (i.e. after applying a NLA and the activation). You're not going to get an exact match between the readout of the ESN to the training data due to the use of penalties (and perhaps an underparameterized or poorly parameterized network).

[MartinuzziFrancesco]

ESNfitted is now merged into main and will be available in ReservoirComputing 0.6.3. Sorry again for my initial confusion, and thanks for the interesting idea for an addition to the package!

I am going to leave this issue open for the moment, since the other ESN variations could use a dispatch for this function, that for the moment has not been added

[sdwfrost]

Thanks @MartinuzziFrancesco. It seems to work well; I've updated my sine-curve gist to incorporate the fitted output - please feel free to use this example if it helps.