Perceptron – Single layer neural network

[Gilles]

Hi Apprentice,
I have implemented the "sigmoid" function, backpropagation with gradient calculation, weight updates, and the displayed result on the screen is: MLP: N/A, and error: N/A.

I don't see where my mistake is.
Can you help me?

Attached are my diagram and my code.

Thank you very much :)

[Apprentice]

We have added your request to the development list.
Development reference 751.

[Apprentice]

Since I don't know what your goal is, I might be completely off.
These are some of my observations.

The main problems.
1. missing Internal Stream
2. two update functions
3. inadequate verification of the number of elements required for calculation
4. Unnecessary loop
5. Call of MLP:update(mode)
MLP external indicator call is not defined.
Note Output stream must not have the same variable name as the variable for the external indicator call.

MLP_VAMA.lua

(10.12 KiB) Downloaded 23 times

Similarly for Multilayer_Perceptron.lua

Multilayer_Perceptron.lua

(9.71 KiB) Downloaded 20 times

[Apprentice]

Not sure about RSI_MLP_sigmoid

RSI_MLP_sigmoid.lua

(26.68 KiB) Downloaded 25 times

Can you review it?

[Gilles]

Hi Apprentice,
At the beginning, I have 4 inputs (RSI) that should allow the model to get a general idea of the market's evolution. The weights of the inputs are used to calculate the output of the first layer.
If you look at my diagram, you'll see that all the connections have a weight. And it's the path they take that allows information to be communicated through the network to the output layer.
It's a forward propagation through the network to generate an output prediction, what we call a pass, a full journey through the network, it's an epoch.
At the end of each epoch, I calculate the error.
So, at this point, the model is not training and not learning.
It's only the implementation of the backpropagation that allows adjusting the network's weights based on the error between the prediction and the actual value.

What allows the model to learn is the adjustment of the weights through the process of forward propagation and backpropagation for a defined number of epochs!
I hope this will help you understand what ++

Note: I don't understand why the loop seems unnecessary to you.

My goal is for the model to learn, and for that, it needs to adjust its weights.

What allows or doesn't allow us to see the decision the model makes is the display of the arrows, conditioned by the slope of the MLP[period] curve.

[Apprentice]

RSI_MLP_sigmoid.lua

(28.49 KiB) Downloaded 22 times

I tried to simplify the structure.
Looks like the weight of the update is problematic.

[Gilles]

Hi Apprentice,
Thank you for your help; your approach is very intriguing, especially with the streams for each weight. This allows checking if the weights are being updated properly. I'll go over all of this with a clear head and get back to you. Thank you very much.