Evolutionary Driving Simulator
2D simulation environment using evolutionary neural networks to train autonomous driving agents without backpropagation.
ML Engineer
Built a simulation framework for training neural networks with genetic algorithms and real-time visualization.
This project trains neural networks with a genetic algorithm instead of backpropagation, which allows experiments in environments where gradients are unavailable or hard to define. The system is a 2D top-down driving simulator where agents navigate custom tracks using sensor inputs. Neural network weights are evolved using a genetic algorithm, with each generation evaluated based on driving performance and improved through selection and mutation.
The architecture cleanly separates simulation, evolution, and visualization. A custom map editor allows rapid environment design, while a real-time viewer shows agent behavior during training. To stop agents overfitting to one track, training ran across multiple tracks with a strategy that encourages them to retain useful behaviors.
Contribution
Designed and implemented the full simulation environment, neural network representation, and genetic training pipeline, including a custom map editor and real-time visualization system.