A Genetic Algorithm which learns how to play google dinosaur game. This is implemented by Yannick. The game uses multiple neural networks specifically for each genome to get high score.

The required modules are

• numpy
• pillow
• pyUserInput
• pyscreenshot

The modules can be installed using the command

pip install numpy pyUserInput pyscreenshot pillow

or by executing the requirement.txt

pip install -r requirement.txt

For creating an environment in anaconda, the command conda env create -f environment.yml should be executed in the anaconda prompt where environment.yml is the environment file.

If there are any issues in installing pykeyboard, the module can be installed manually by running python setup.py install in command line after going into the directory ./Manual/PyUserInput. This problem may be caused by pyHook as it is not available for newer versions of python.

The System Modules consists of three main parts. They are:

Each module has a set of operations in which some of them requires data from other modules. The given architecture diagram illustrates about these kinds of operations.

• Generation
  • The Generation module is the main module used in performing selection and mutation process of best genomes in the population of genomes.
  • The Genetic Algorithm is executed in this module.
• Network
  • The Network module is used for creating and updating the neural network for each genome.
  • Neural Network is created in this modules.
  • Backpropagation Algorithm is executed in this module.
• Scanner
  • The Scanner module helps the system to view the game.
  • It is used for calculating the distance, time, speed and length of obstacles.

The Procedure can be illustrated using the following diagram:

• Machine does not handle velocity well enough. In this, we notice velocity makes great impact over the jumping position selection.
• The dinosaur colour in the game changes which creates a problem for scanning the obstacles and also the software can’t find the game.
• Sometimes the dinosaur does not consistently score high due to random frame drops that occur while learning on a CPU only system. The current model was trained only on CPU for which some game features were stripped. GPU training would give a consistent frame rate while accommodating more features.