Artificial Neural Networks working with Hal

Hi,

I was curious how to use Artificial Networks. I found a nice blog how to do it.
This blog explain's a basic Xor logic example. ( 2 inputs, 1 output )
In linuxcnc the component xor2.comp is more or less the same.

I compiled the c++ example and it worked. Output is oke.
Now i am planning to test this Artificial Network working within a hal component.

If the program is trained, it can predict a solution.

What should be a nice thing to predict in linuxcnc hal?










 

backlash compensation.
actually any compensation.

PID control?

torch height (voltage)
torch void detection (voltage spike outside of standard deviation)

Hi,

I just finished the hal implementation of the arteficial neural network performing a xor2 function.
The code is at github, you can look at it. It works ok with linuxcnc on the servo-thread.

First i train the ai providing the training_data file. # setp train 1
Then i run the ai with # setp run 1
Once the ai is trained. You don't have to train it again next time linuxcnc starts.



I think when to make any implementation request's succesful we have to consider the following:
- the ai has to be trained. User has to provide a training set.

The training set can be for example this file , where line 2,3 are a set. line 4,5 are a set, etc.
Line 2 = input 0,1 (xor2 input)
Line 3 = result input 0,1 (xor2 output)
and so on.

I see the request's are manly for position control purposes.
 

??
I am confused.... I've read... I don't understand how it works (yet!)... off to read again....

Hi,

The library that is used.
The library set's up the neural network, input's, output's, layers between etc.

In the example there are 2 input's and 1 output.
The ai is trained from a data sheet how the result on the output should be, when the 2 input's are at different states.

Normally we should code this in c or c++ by :
if(input0==0 && input1==0){output=1;}
if(input0==1 && input1==1){output=1;}
if(input0==0 && input1==1){output=0;}
if(input0==1 && input1==0){output=0;}

In the "ai network" we can train above situation by giving the desired output like :
input0=0 input1=0
-> ouput=1
input0=1 input1=1
-> ouput=1
input0=0 input1=1
-> ouput=0
input0=0 input1=1
-> ouput=0

Behind the screen the ai is trained. This example is trained within 1 second. But i have already experienced a training period
of around 1 minute for a more complex situation.

You can in fact cheet the output with an ai quite easy by giving wrong results in the training sheet. For example we now
have cheeted the ouput to a inverted xor2 output.
input0=0 input1=0
-> ouput=0
input0=1 input1=1
-> ouput=0
input0=0 input1=1
-> ouput=1
input0=1 input1=0
-> ouput=1

I hope this tiny example helps a bit.