Thıs repository is documented and cleaned version of my old project I did back in 2019 which can be found here and it is a toy project for me to learn about neural networks and evolutionary algorithms
The Project is about the implementation of deep neuralevolution algorithm and driving autonomous agents on different types of parkours. Agents have laserscan sensors that casts rays around its surrounding and get distance mesaruments. These distance values and agents velocity and orientation information is given to the network. As an output we get acceleration and steering changes that agents need to do As the reward system checkpoints(blue dots) are used
- Neural networks and matrix calculations from scratch
- Genetic algorithm implementations
- Custom simulation environment
- Map generation tool for creating unique parkours
You can run the simulation with.
python3 main.pyThe script uses "/mapData.json" for getting map information and "/bestcar.json" for getting and saving model parameters
There are several example parkours in "/parkours" folder. In order to use them, you can carry the desider json file to project directory and rename the file to "mapData.json"
Similarly you can use pretrained model in "/train models" directory, just rename it to "bestcar.json"
Aside of already generated maps you can also create and edit your own custom maps to train/test your models.
To run the map generation tool
python3 MapTools/mapMain.pyWith that editor you can place walls, determine start/finish point and adding checkpoints. The map is going to saved when you exit the application
Maps you created are saved in "/mapData.json" file, dont forget to extract that file
If you want to create an empty map from scratch you can replace "/mapData.json" file with "MapTools/emptyMap.json"
Note : Checkpoints cant be deleted and you need to place them in order from start to finish of the parkour