First of all, I'm not a machine learning student and I've got no experience in this domain, this is just a test experience and I'm shit at theory so if you're a machine learning worker I sincerely excuse for the eye bleeding

This project is a basic game played by a simple neuronal-based learning algorithm

The goal of the game is for the player ( blue ) to reach the treasure ( yellow ) while dodging the enemies ( red ). I initially wanted to put walls so that would make more obstacles on top of enemies but they play the same role as enemies as you need to dodge them.

The program itself is using 2 layers of 4 neurons to make its decisions ( I'm still not sure if we can call that layers tho so let's just call them neurons groups ).

The first neuron groups ( we'll call them the case neurons ) is in charge of the cases close to the player ( there is one neuron for each case close to the player: top, right, bottom and left )

The second neuron groups ( we'll call them the moves neurons ) is in charge of the direction the treasure is compared to the player ( top, right, bottom and left too )

Each of the 8 neurons, given an input value, outputs a value ( between 0 and 1 for the case neurons and between 0 and 2 for the moves neurons ) corresponding to the worth of the action

For example if we ask the neuron in charge of the case top from the player and there is an enemy in it he'll answer 0 because the action has no value ( we die if we go in the case ), if there is the treasure in it he'll output 1 ( we win if we go to that case, the best case possible ), if there's nothing we neither die nor win so he'll output 0.5

The second neurons groups will output a value > 1 if we get closer to the treasure ( if we go right while the treasure if on our right ) and a value < 1 if we get further from the treasure

The algorithm navigates through the map while taking the best actions possible ( dodging enemies while getting closer to the treasure ) until it reaches it and win

So, how is this machine learning and how does it work ?

Each of the neurons action-values are registered in a JSON array corresponding to each neuron like this:

`{id: The neuron's id, empty: Value if the case is empty, enemy: Value if there is an enemy, treasure: Value if there is the treasure}'

At the beginning of the program and as you can see in the file data.js, all values are initialized to -1 ( which means the value is unknown). Before taking decisions and doing the best action possible, the algorithm looks if one of the action equals -1, and if there is one it does it and looks for what happens: that's called the reward

To simplify it: if there is an enemy is the case and the value is -1, the program doesn't know what happens so it just goes in and dies. Then it saves the value as 0 because if's the worst action to do. That way, the program learned itself what happened when you go right into an enemy, and it just pretty luch learns everything like that

That's why, when you launch it, if just suicides into the enemy 4 times and then dodges it everytime after

If you have knowlegde on the subject and have a comment to make about this project don't hesitate, I was just doing this to try a little bit but I would love to learn more about it and learn the good practices of the field :)