Machine learning for autonomous driving

Ideja projekta bila je pokušati uz pomoć algoritma za strojno učenje naučiti računalo samostalno upravljati automobilom na stazi. Grafičko sučelje je riješeno uz pomoć pygleta, moćne python knjižnice za grafiku. Kretnje automobila te računanje kolizije riješeno je korištenjem vektorske geometrije i trigonometrije 2D svijeta. Implementirao sam neuronsku mrežu kao sučelje za računalo preko kojeg može samostalno upravljati automobilom. Ulazne vrijednosti za neuronsku mrežu su bili pet senzora na automobilu koji su javljali automobilu njegovu udaljenost od rubova staze u pet različitih smjerova. Osim pet ulaza, neuronska mreža se još sastoji od dva skrivena i jednim izlaznim slojem. Sve zajedno sačinjena od točno 100 neurona. Kako bi neuronska mreža naučila upravljati automobilom korišten je genetski algoritam. Evolucijski algoritam stohastičke prirode, koji je odličan za rješavanje težih problema optimizacije. Kao rezultat projekta dobivena je aplikacija koja demonstrira učenje neuronske mreže primjenom genetskog algoritma na primjeru vožnje automobila po zadanoj stazi. Naime lakše staze je algoritam uspješno odvozio, no kada se staza zakomplicira izgleda da algoritam zapne na nekom od lokalnih minimuma, te ne uspijeva završiti stazu.The idea behind the project was to use a machine-learning algorithm to teach a computer to operate a car on the track on its own. The graphical interface is solved with the help of pyglet, a powerful python library for graphics. Car movements and collision calculations were solved using vector geometry and 2D trigonometry. I have implemented a neural network as an interface for a computer through which it can operate the car independently. The input values for the neural network were five sensors on the car that reported to the car its distance from the track edges in five different directions. In addition to the five inputs, the neural network also consists of two hidden and one output layer. All together made up of exactly 100 neurons. A genetic algorithm was used to help the neural network learn how to operate the car. GA is an evolutionary algorithm of a stochastic nature, which is great for solving more difficult optimization problems. As a result of the project, an application was obtained to demonstrate the learning of the neural network by applying the genetic algorithm to the example of driving a car along a given track. Namely, the algorithm successfully drove the easier paths, but when the path becomes complicated, the algorithm seems to get stuck at one of the local minimums and fails to complete the path