NSS block in Simulink with Swish Activation Fcn

Hi,

I have trained my neural state space with Swish activation function using MATLAB 2024b version. However I have not been able to test the accuracy using the Neural SS model in Simulink. I have multiple warnings and thereafter followed by error. If anyone can help, please kindly let me know. Thank you.

Warning: nnet.cnn.layer.SwishLayer is not supported. A placeholder subsystem will be added for 'act1'.

Warning: nnet.cnn.layer.SwishLayer is not supported. A placeholder subsystem will be added for 'act2'.

Warning: nnet.cnn.layer.SwishLayer is not supported. A placeholder subsystem will be added for 'act3'.

Warning: nnet.cnn.layer.SwishLayer is not supported. A placeholder subsystem will be added for 'act4'.

Warning: nnet.cnn.layer.SwishLayer is not supported. A placeholder subsystem will be added for 'act5'.

Warning: 'Input Port 1' of 'nlss_simulator/Neural State Space Model/NSS/State Network/State

Network/NetworkSubsystem/act4/out' is not connected.

Warning: 'Input Port 1' of 'nlss_simulator/Neural State Space Model/NSS/State Network/State

Network/NetworkSubsystem/act2/out' is not connected.

Warning: 'Input Port 1' of 'nlss_simulator/Neural State Space Model/NSS/State Network/State

Network/NetworkSubsystem/act1/out' is not connected.

Warning: 'Input Port 1' of 'nlss_simulator/Neural State Space Model/NSS/State Network/State

Network/NetworkSubsystem/act5/out' is not connected.

Warning: 'Input Port 1' of 'nlss_simulator/Neural State Space Model/NSS/State Network/State

Network/NetworkSubsystem/act3/out' is not connected.

Error using test (line 19)

Compilation of model 'nlss_simulator' failed while trying to resolve underspecified signal dimensions.

Caused by:

Error using test (line 19)

Matrix multiply dimensions propagation error. Error occurred while setting 'Input Port 2' of

'nlss_simulator/Neural State Space Model/NSS/State Network/State Network/NetworkSubsystem/dxdt/Matrix

Multiply' to have the dimensions [1 x 1]. A possible cause for this error is that these dimensions do not

agree with the partial dimensions information present on other ports

Error using test (line 19)

Error in port widths or dimensions. 'Output Port 1' of 'nlss_simulator/Neural State Space Model/NSS/State

Network/State Network/NetworkSubsystem/dxdt/Reshape' is a [1x1] matrix.

Answer 1

[Tianyu]

Hi Saskia,

As warning suggests, swish is currently not supported in Simulink.

We have fully supported the activations in command line, but for Simulink, elu, swish, softplus and scaling are not yet added.

We should have mentioned this in doc. Sorry about the confusion.

For now, please use other activations if simulation is needed.

[Saskia Putri]

I understand. Thank you for the clarification.

Answer 2

[Subhajyoti]

Hi @Saskia Putri,

It is my understanding that you are trying to test the Neural State-Space Model that you have trained. The error you are encountering, “Multiply' to have the dimensions [1 x 1]. A possible cause for this error is that these dimensions do not agree with the partial dimensions information present on other ports)” indicates a mismatch between the input and expected shapes in the model, which is generally sensitive to input dimensions and layer compatibility.

Here, in the following implementation, I have estimated and validated a neural state space model with “Swish” as activation functions, using dummy values.

rng(0)

Ts = 0.1;
sys = ss(c2d(tf(1,[1 1]),Ts));

sys.b = sys.b*sys.c;
sys.c = 1;

te = 0:Ts:10; 
ue = randn(length(te),size(sys.B,2));

ye = lsim(sys,ue,te,zeros(size(sys.B,1),1));

tv = 0:Ts:1;
uv = randn(length(tv),size(sys.B,2));

yv = lsim(sys,uv,tv,zeros(size(sys.B,1),1));

% Create object for time series data
estimationData = iddata(ye,ue,0.01);
validationData = iddata(yv,uv,0.01);

% Construct neural state space object
nss = idNeuralStateSpace(1, NumInputs = 1, Ts = 0.01);

% Specify state network parameters
stateNetwork = createMLPNetwork(nss, 'state', Activations = 'swish', LayerSizes = 16);
nss.StateNetwork = stateNetwork;

% Specify training algorithm parameters
opt = nssTrainingOptions('adam');
opt.MiniBatchSize = 100;

% Estimate neural state space model
sys2 = nlssest(estimationData,nss,opt);

% Display results
compare(estimationData,sys2);
clear estimationData nss opt stateNetwork;
title('Fit to estimation data');

figure;
compare(validationData,sys2);
clear validationData;
title('Fit to validation data');

Additionally, you can refer to the following resources to know more ‘Neural State-Space Models’ in MATLAB:

• https://www.mathworks.com/help/ident/neural-state-space-models.html
• https://www.mathworks.com/help/ident/ref/estimateneuralstatespacemodel.html

[Saskia Putri]

This is very helpful. thank you.