Chrome Dino Game AI

A Python implementation of the Chrome Dinosaur Game with NEAT (NeuroEvolution of Augmenting Topologies) algorithm that learns to play through population evolution.

Overview

This project recreates the Chrome browser's dinosaur game and implements a neural network-based AI that learns to play the game through genetic evolution. The dinosaurs learn to jump over cacti and duck under birds to achieve higher scores.

Repository Structure

• main.py: Primary game application with visual interface showing dinosaurs learning in real-time
• train.py: Headless version of the game for faster training without rendering
• settings.py: Game configuration and parameters
• requirements.txt: Project dependencies

Game Module

• game/dinosaur.py: Dinosaur class with movement and collision detection
• game/dinosaur_controller.py: Interface between the neural network and dinosaur actions
• game/entities.py: Game entities (obstacles, backgrounds, etc.)
• game/population_controller.py: Manages dinosaur population and evolution process

NEAT Implementation

• neat/ffn.py: Feed-forward neural network implementation
• neat/genome.py: Genome class for storing network structure and weights
• neat/evolutionary_operators.py: Mutation and crossover operations
• neat/activations.py: Neural network activation functions
• neat/edge.py, neat/node.py, neat/counter.py: Supporting classes for network structure

Utilities

• utils/serialization.py: Functions for saving and loading trained populations
• utils/network_visualizer.py: Tools for visualizing neural networks

Other Directories

• assets/: Contains game sprites and images

Installation

1. Clone the repository
2. Install dependencies:

   pip install -r requirements.txt
   

Usage

Playing with visual interface

To run the game with visual interface showing the dinosaurs learning in real-time:

python main.py

Training without visualization

For faster training without graphical rendering:

python train.py

This headless mode saves training progress to TensorBoard logs and periodically saves the population to population.json.

Monitoring training progress

To monitor training metrics with TensorBoard:

tensorboard --logdir=logs