How to update the weights of a Shallow neural network by supplying one sample at a time?

I would like to use a shallow neural network inside a simulation loop. Every loop will generate a new input that is used to update the weights of the neural network. This online training of a shallow network.

To clarify, I am not asking how to use concurrent nerual network, or reinforcement learning. I am just asking how to make the shallow neural network perform one backprobagation operation based on one desired output. Nothing more. Thanks.

[Emmanouil Tzorakoleftherakis]

By the way, if you are willing to try a packaged solution, please take a look at this example that shows how to train a neural state space model.

Answer 1

[Venu]

Hi @EngM,

You can refer to the documentation below to start with.

https://www.mathworks.com/help/deeplearning/ug/multilayer-neural-networks-and-backpropagation-training.html

To perform online training of a shallow neural network in MATLAB by supplying one sample at a time and updating the weights based on that sample, you can use the "adapt" function.

The "adapt" function is designed to perform online updates to the network weights and biases according to the specified training function (example "traingd") for each new input-target pair.

Read the "adapt" documentation for more understanding.

help adapt 
doc adapt

Hope this helps

[EngM]

Hello, thanks for answering, I tried the adapt function, the weights/biases don't seem to be adapting (getwb(net)) .

I am using ode23 to simulate a simple 2-dof linear system that the NN is trying to fit, please review. Thanks..

The ode function file is below: input is sin(3*t)., the net is passed on with the ode function

function [dxdt] = SysLTI_2dof_NN(t,x,net)

% Auto generated function on 04-Jan-2024 11:19:18

dxdt = zeros(4,1);

dxdt(1) = x(3);

dxdt(2) = x(4);

dxdt(3) = 50*x(2) - 150*x(1) + sin(3*t);

dxdt(4) = 100*x(1) - 100*x(2) - x(4)/5;

% get the output

y1 = x(1);

y2 = x(2);

net = adapt(net,sin(3*t),[y1;y2]); % adapt function is called here.

end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% network shallow feedforward:

net = feedforwardnet(10,'trainlm');

net = configure(net,input,output);

% perform initial off line training so network can get inputs/outputs configuration

[net,tr] = train(net,input,output); % input/output data were provided.

% online training

tspan = [0 1]; % simulation time.

x0 = [0;0;4;0]; % initial point of the ode.

%options = odeset('OutputFcn',@adapt_net);

[t,x] = ode23(@(t,x)SysLTI_2dof_NN(t,x,net),tspan,x0);