🏓 Pong AI Trainer

A Python implementation of the classic Pong game featuring a custom-built Neural Network built from scratch (no ML libraries like TensorFlow or PyTorch).

The AI learns to play in real-time using Supervised Imitation Learning, training itself to mimic a perfect heuristic algorithm.

📋 Features

• Custom Neural Network: A manual implementation of a Multi-Layer Perceptron (MLP) using only NumPy.
• Real-time Training: Toggle between "Training Mode" (accelerated speed) and "Player Mode" (human vs AI) instantly.
• Visualization: See the AI improve as it iterates through thousands of game loops in seconds.
• Clean Architecture: Separates game logic, rendering, and the neural network class.

🧠 How It Works

The Model

The AI uses a Feedforward Neural Network (Perceptron) with the following topology:

• Input Layer (3 Neurons): Normalized values for Opponent Y, Ball X, and Ball Y.
• Hidden Layer (5 Neurons): Uses the Sigmoid activation function.
• Output Layer (1 Neuron): Determines direction (Value > 0.5 = Down, < 0.5 = Up).

The Learning Process

Unlike Reinforcement Learning (which learns from rewards), this agent uses Supervised Learning.

1. The "Teacher": A simple algorithm calculates the perfect move for every frame (aligning the paddle center with the ball).
2. Backpropagation: The network compares its own prediction against the "Teacher's" move.
3. Optimization: It calculates the error and adjusts weights using Gradient Descent.

🚀 Getting Started

Prerequisites

You need Python installed along with pygame and numpy.

pip install pygame numpy

### 🎮 Controls & Game Modes

The game has two distinct modes. You can toggle between them at any time.

| Key | Function |
| :--- | :--- |
| **SPACE BAR** | **Toggle Modes** (Switch between Training and Playing) |
| **UP Arrow** | Move Player Up (Player Mode only) |
| **DOWN Arrow** | Move Player Down (Player Mode only) |