Overview
Pong AI was a university project for an Introduction to Artificial Intelligence class. The assignment was to train a model that would improve iteratively and eventually learn to play a game of our choice.
We chose Pong partly because of its history. It is one of the earliest and most recognizable computer games, and its rules are simple enough that the AI part could stay as the main focus.
The project is a small Python implementation built around Pygame and NEAT. It has two paddles, a ball, scoring, collisions, and a training loop where genomes play against each other across generations.
What It Does
The code contains a simple Pong implementation with separate modules for the ball, paddles, and game loop. On top of that, there is a NEAT training runner that evaluates pairs of genomes against each other.
Each neural network receives a compact game state:
- the paddle position
- the horizontal distance to the ball
- the ball height
From those inputs, the network chooses whether to stay still, move up, or move down. Invalid paddle moves are penalized, and fitness is based mainly on survival time and successful hits.
The repo includes saved checkpoints and a best.pickle model from training.
After training, we also tried to beat the last generation ourselves, which was probably the most entertaining part of the project.
Why NEAT
NEAT is not a modern or trendy AI approach. It is an older evolutionary algorithm for evolving neural network topologies and weights over generations.
That made it a reasonable fit for this assignment. We did not need a large reinforcement learning setup or a heavy ML pipeline. We needed something that could be wired into a game loop without turning the whole project into overengineering.
The NEAT config was initially generated with help from an LLM, because at that point we were still figuring out how to approach training in practice. After that, we used it as a starting point and mostly focused on making the training loop work.
Context
At that stage of university, I honestly thought requiring every group to write its own game was overkill. Most groups were still learning how to write correct code comfortably, so I expected many people to copy the game part from a tutorial anyway.
I had an easier time because I already had four years of technical school experience behind me, and some of my friends were in a similar place. That made the game implementation less intimidating, but the training part was still new to us.
Looking back, this is a very early AI project. It is not especially advanced, but it was a useful first contact with training something that changes its behavior over time instead of just following hand-written rules.
Takeaway
Pong AI was a small class project that ended up being more fun than it probably should have been. It was a bit overkill for the course level, a bit improvised, and one of my first practical AI tasks.
The nicest part was simple: after training, there was actually something on screen that could play back.