Developing a classic pong game with a minimalistic 2.5D graphics in Unity. Implementation of nural networks to make an A.I pong agent. Human vs AI , AI vs AI mode. training , tuning the AI agent for smart gameplay style
gameplay image: https://github.com/parikshitsaikia1619/AI_agent_playing_PONG/blob/master/pics/Screenshot%20(51).png?raw=true
It's damn simple.., just add lots of static meshes, make it look a PONG court , tag the walls for collsion detection and raycasting. Add materials to make it look beautiful. Tadaaa...it's done.
you can do it man , I believe in you
here it should look like: https://github.com/parikshitsaikia1619/AI_agent_playing_PONG/blob/master/pics/Screenshot%20(47).png?raw=true
I have added rigidbody to both of them , which makes movable by add force, velocity ... you know by Physics So, for the paddle I have created a scripted like :
if (key W is pressed ) add constant velocity of say x (in up direction) in paddle. if (... S is ..pressed) add constant velocity -x
Piece of cake right.. let move on to ball
I have added a physics material on ball , for its bouciness. its simple , on starting I added a random velocity to ball ,(rigidbody of ball , gravity is set to 0) random initial velocity , zero gravity ... keep moving forward ... bouncing forever . Simple Logic isn't it :)
I have done some collision checks on the ball , to know what it is hitting... you know DEBUGGING, and also to change color of the ball , updating score in UI.
polishing the game :)
----- OUR GAME ENVIRONMENT IS DONE ,NOW IT'S TIME FOR AI------------
Now it is the boss fight level
TIME TO USE MACHINE LEARNING * sigh * , I will keep as simple as possible So, I written my ANN from scratch. for this game we will use our ANN.
So, lets figure out our inputs and outputs:
inputs: 1. ball position x
2. ball position y
3. ball velocity x
4. ball velocity y
5. paddle position x
6. paddle position y
outputs: paddle y velocity
So, we created our 6 input , 1 output , 1 hidden layer with 4 nodes , learning rate 0.11
hidden layer activation function is relu , for output act. func. is tanh ( for -1 to 1 velocity)
now for the interesting part creating error function...OK not that interesting, let keep moving
for the error function we will give the difference between the ball's y position (hitting on the backwall when misses) and paddle's y position.
we can get the ball's hitting position by RAYCASTING
that sounds terrifying
this screenshot will give you an idea: https://github.com/parikshitsaikia1619/AI_agent_playing_PONG/blob/master/pics/Screenshot%20(52).png?raw=true
So basically our model will take the inputs , give the output , minimise the error by updating the weights and bias to get better
typical forward and backpropagation
FINALLY.... our model is ready our output will use to move the paddle.
READ THE SCRIPTS FOR DETAILED UNDERSTANDING..
To play the game just go to Pvs AI folder , select the exe file