Matchbox neural network

I do not remember where I took this information, but it reflects the essence of neural networks best of all. On fingers.



Rules of the game. NA learns to play the game "11 sticks". You can take either 1 wand or 2. You need to pull out all the wands last.



We take 10 matchboxes and put buttons of two colors in each. For example, black and white. One by one. The number on the box will be responsible for the number of sticks currently in use.

')

For example, the NA starts the move.



1. Close our eyes and randomly pull out a button from box number 11. If black, then take one stick, if white - two. (Let it be white - 2 sticks).

2. The progress of the person. For example, he took 2 sticks.

3. Next, take the box number 11-2-2 = 7 and at random take out a button from it.



So until the game ends.



If the National Assembly won, then we encourage the solution found by adding one button at the top of the boxes to the same color as it was drawing. If the NA has lost, then we punish by removing the box of the elongated button from the last one.



That's the whole neural network of 10 nodes which, initially, not even knowing the rules, learns to play and starts to beat the person. If you change the rules and, for example, the one who takes the last sticks loses, then the National Assembly will retrain and start winning again.



Here, of course, the scale is insignificant, but it well shows that the National Assembly is good because there is the possibility of flexible learning and adjusting to the rules of the game.