Neural Network Demo

Описание: Github: https://github.com/matthew-ch-robbins...

About:
The simulation is based on a neural network to guide the car around the track and a genetic algorithm used to pick and breed the most successful cars of each generation.

For the people interested in how it works:
The car uses its 5 feelers to calculate a normalised intersection depth with the track's edges and then feeds these 5 values as inputs to the neural network. The inputs are then passed through a hidden layer of 8 neurons and finally to an output layer of 2 neurons: a left and right steering force. These forces are used to drive the car forward and turn the car.

Each car represents a different genome in a generation (or a unique set weights for the neural net) which are evaluated and potentially carried through to the next generation by a fitness score. The fitness score is based on the distance travelled along the track as well as bonus points for hitting checkpoints (this encourages the network to navigate around tricky sections of the track).

In the video, the neural network / genetic algorithm finds a solution at about 2.25. FRAPs interfered a little by clamping the framerate, so it took a bit longer than usual for it to find a solution.

Legend:
The green dots are the intersection points of the feelers. The blue lines are the feelers
The orange highlighted track geometry are sections being considered for collision.
The blue circle is the radius used to search for track nearby track geometry.