This project was the product of a competition between a few of my peers to produce AIs for pong and see who could make the best (make them vs each other). Unfortunately, I was the only person to produce a final product so I cannot say whether mine performed the best (although I guess it did win by default?). My methodology and techniques used are extremely rudimentary and were based on extremely limited research, hence, there is significant room for improvement.
The project is made using Python and the following libraries:
- Numpy
- Pygame
- Math
- Random
I feel as though this project taught me a lot, however, much of it was basics of project management, planning and basic coding principles rather than AI building. It was a lot of fun to make and feel free to give it a test. Simply run the "ModelShowase.py" file, you can also train up a new AI and watch the training process using the "main.py" file (uncomment some of the code to watch the process).
The rules of the competition allowed for access to your paddle y-coordinate, the ball coordinates (x,y) and the balls velocity vector (x,y). Hence, the model will take in 5 inputs. My goal was to create an AI that would take in these 5 inputs and output a singular boolean value (1 go up, 0 go down). Hence, using my extremely flawed and gap-filled understanding of what an AI is and how it works I produced a model.
My model consists of an array of N rows and 5 columns with N being the number of 'hidden layers' (or at least that's what I am referring to them as) this array holds all the weights which are initialised as a bunch of random numbers, there is then an additional variable "Bias". As any function can (to a degree) be represented by a polynomial function and an AI is a complex function that can convert these inputs into the desired output:
I decided that I would sum each weight * the input value ^ the row (hidden layer), and then add the bias. The result of this sum would then be thrown into a sigmoid function to produce a value between 0 and 1 and then rounded in order to acquire the boolean output that is desired.
In order to train the model and improve it I construct an array of models and simulate hundrends of games of pong taking note of how long the games lasted. The games have another paddle that has its y-coordinate set to that of the ball. Once all the games end I find the model that achieved the best result and populate the other models with its weights +- a random amount. This random amount shrinks as the epoch increases in order to refine the model further as time progresses. This makes the evaluation of the model not care whether they win/lose but rather just how long the game lasts. Once training is completed the weights and bias are then saved into a txt file in order to be loaded later.