7

Make your Arduino learn

TUTORIAL

bad. The process involves finding out how many of the

user data cases a combination would have predicted

correctly. For example, here are the prediction scores

for three random variable value combinations :

Flip_flag : true, thresh : 500

100 – on wrong

300 – on wrong

600 – off correct

800 – off wrong

850 – off wrong

Total correct = 1

Flip_flag : false, thresh : 200

100 – off correct

300 – on wrong

600 – on wrong

800 – on correct

850 – on correct

Total correct = 3

Flip_flag : false, thresh : 700

100 – off correct

300 – off correct

600 – off correct

800 – on correct

850 – on correct

Total correct = 5

Above

You can do this

using either the

Arduino IDE or

create.arduino.cc

MORE ADVANCED LEARNING

Usually, machine learning systems have far more

than two values to optimise (there can be millions

of values in advanced image-detection systems). To

train those systems it’s not feasible to just try every

combination, so instead they’ll use something called

gradient descent, which is a lot more efficient but

more complex to implement.

Now imagine we scored the performance of every

combination of threshold and flip_flag; that’s roughly

2000 different combinations. The Arduino does it in

a blink of an eye. Here is the Arduino code that loops

through each combination of variable values:

void    optimise(){

        //  do  brute   force   search  to  optimial    coef    and

intercept   values

        int best_score  =   0;

        int current_score   =   0;

        for (int    flip_flag   =   0;  flip_flag   <=  1;  flip_flag++){

                for (int    thresh  =   0;  thresh  <   1000;   thresh  +=

30){

                        current_score   =   score(thresh,   flip_flag);

                        if  (current_score  >   best_score){

                                learnt_thresh   =   thresh;

                                learnt_flip_flag    =   flip_flag;

                                best_score  =   current_score;

}

}

}

}

You can see that we score the performance of each

value combination and save the value combination that

performs best. You may think of this as automated

‘trial and error’, that’s because machine learning is

effectively trial and error – but isn’t human learning

also trial and error? Try something, evaluate your

performance, change what you’re doing, repeat!

To score the performance, we are counting

how many of the saved user data points the value

combination correctly predicts. So if we only stored

three data points, the best possible score a value

combination can get is three.

DATA COLLECTION

We want this system to learn based on the most

recent user input only. This will allow the system to

quickly adapt to new user behaviour. We can achieve

this by just storing the last ten times the user toggled