Automating Chrome Dinosaur using NEAT

After watching Code Bullet's video about automating the Dinosaur game on Chrome ( when you have no connection ) using NEAT (Neuroevolution of augmenting topologies) algorithm. I had no idea how it worked so this project was the perfect moment to discover this algorithm. I used python and two libraries helped me : pygame and neat-python . You can download the project and the libraries by running in your terminal :

git clone [email protected]:Brice-Vergnou/Smart-Dino-NEAT.git
pip install -r requirements.txt

Then you just have to run main.py in the Scripts folder. Make sure you have Python 3 installed.

It's a tweaked version of my version of the game suited to the needs of the algorithm

The NEAT algorithm is basically natural selection. You pick the best dinosaur of the generation and use a slightly modified version of him the next generation. To help this dinosaur , you just have to tell the inputs ( what he sees ) and the output ( what he does ) , making a neural network.

It is how we select the best dinosaur. The higher it is , the better. In our case , it can be modified by three facts :

• If the dinosaur survives ( slightly increases the fitness function )
• If an obstacle is now behind the dinosaur ( increases the fitness function )
• If the dinosaur hits an obstacle ( decreases the fitness function )

There are 8 inputs , I tried to replicate how a human think when they play the game :

• The y position of the dino
• The distance between the dino and the obstacle
• The y position of the obstacle
• The speed
• The width and the height of the obstacle
• The distance between the bottom of the dino's hitbox height and the obstacle's one ( it helps with birds )
• The type of obstacle ( Bird or cactus )

There are only two outputs :

• Ducking
• Jumping

As I don't have a lot of compute power , I didn't use any hidden layer , so there are only the 8 input nodes and the 2 output nodes. After each generation , the weights between these nodes are adjusted. These weights are basically how a specific input influences the output ( e.g. If the y position is equal to the y position of the dinosaur , the dinosaur should jump )

Even though the algorithm sometimes struggle to evolve ( the dinosaurs sometimes keep ducking or jumping but it's still rare ) , I'm pretty satisfied as a first NEAT project :)