#Pong with neural nets

I want to study and further understand neural nets and evolutionary algorithms. For this reason i decided to rewrite my old PoC pong clone into using neural net "brains" as players and pit them against a bad hardcoded AI.

##What have i done

• Written small clone of pong This is a bad implementation, should probably be rewritten

• Written Neural Net
  This is based on 'neural nets demystified' https://www.youtube.com/watch?v=bxe2T-V8XRs

Merged neural net with pong to control player
The neuralnet player is initialized with random weights Takes ball coordinates as input and outputs a value that is interpreted as an action. The action is either move up, move down or stay.

###What do i want to do I want to create bat evolution and evolve better bats

####For this i need to

• Model weights (or neuralnets?) as bytes
• write evolution code
  -- Mutation
  -- combining two 'brains'
  -- decide on fitness functions
  -- rewrite pong to play a game, not play forever
  -- Needs to output results, preferably to a textfile
  -- Create 'tournament' code for selecting brains to bring on to next generation

###I also want to track this stuff, preferably using matplotlib ####For this i need to

• Learn matplotlib
• Figure out what data to save

• There is no reason to watch the games, if i want to run for instance 10.000 iterations/generations.

I have a neural network class.
The class has various properties:

#What happens?
When the forward function is called, a matrix of input is supplied.
In the pong game, this matrix is a 2x1 matrix, containing distance from paddleY to ballY and the ball's xCoordinate.

The dot-product of X and W1 is then calculated and stored in z2.
This is run through the activation function (sigmoid), and stored in a2.
The dot-product of a2 and and W2 is then saved in z3.
z3 is run through the activation function and saved in yHat.
yHat is returned.

The sigmoid function is:

def sigmoid(self, z):
    # Apply sigmoid activation function
    return 1 / (1 + np.exp(-z))

This results in a value between 0 and 1.
If the value is between 0.4 and 0.6, the paddle stays where it is.
If it is less than 0.4, it moves up.
If it is more than 0.6 it moves down.

Dependencies

• Pygame
• Numpy