This project utilizes the python implementation of NEAT to play Sega Genesis Frogger.

An evolving artificial neural network attempts to simultanesouly learn weight values and the structure of the neural network that will perform best (via 'rewards' tracked in retro-gym).

(back to top)

On Windows, ensure you install conda and git

The recommended installation method is described below, but if you prefer to install via pip then ensure you have the following packages installed

pip install gym-retro==0.21 matplotlib neat-python opencv-python numpy

Note: During development gym==0.21 was used due to compatibility issues

1. Clone the repo

   git clone https://github.com/rdarneal/neat-retro-frogger.git

2. Navigate to cloned directory

   cd neat-retro-frogger

3. Install conda environment

   conda env create -n retro-env --file environment.yml

4. Actiate the environment

   conda activate retro-env

5. Load the ROM(s) using gym-retro

   python -m retro.import /path/to/your/ROMs/directory

Note: This repo does not contain the ROM file for Frogger-Genesis. Check the No-Intro Collection on Archive.org for ROMs.

(back to top)

To jump right to training

python app/parallel-runner.py

This will launch into training via natural selection. The script will run until an individual meets the fitness threshold 5,000 points OR 100 generations of Frogger, whichever is sooner.

If you want to see the results, run:

python app/playback-winner.py

Which shows how species performed over multiple generations

Along with the average fitness over time, determined by the retro-gym reward function (in this case tied to how well frogger crosses traffic)

By default, each generation includes a population of 100, and random values are passed to the retro-environment in order to control frogger.

Each generation is then divded into 'species', or groups of individuals with similar performance, and a new generation is created.

The .config file specifies the parameters used in training the network - which includes how long species survive stagnation, , check out the neat-python documentation for additional details on what these parameters do.

You can train a new Frogger model using the parallel-runner.py script, you can play with the following parameters:

• In the __main__ script:
  • num_workers, a keyword argument of neat.ParallelEvaluator() can be scaled to the number of CPU cores you would like to use for parallel training.
  • n keyword argument in the p.run() method defines the number of generations that will be used unless the fitness condition is met.

If you want to modify this for other games:

• Each python script needs to be modified to match the retro environment to the game you want to use
• You will need to download and use retro to load the ROM
• The config file is currently setup for Sega Genesis
  • Modify num_inputs to match the screen size of console
  • Modify num_outputs to 12, which should match # of buttons
  • Modify the python scripts to refactor use of buttons

(back to top)

Robert Darneal - python-dev@robertdarneal.com

Project Link: https://github.com/rdarneal/neat-retro-frogger

(back to top)