Racing AI - a machine learning demonstration

(Takes a while to load)

Adjust the parameters of the neural network and its training, select 4 different maps, and race the robot you have trained.

You control the blue car with WASD

Hidden Layer Neurons: controls the number of neurons per hidden layer, with 21 input neurons and 4 output neurons

Hidden Layers: The number of hidden layers the neural network has. The number of neurons is controlled by "Hidden Layer Neurons".

Bots per Generation: Number of cars that race each generation, with higher numbers increasing lag, but also increasing population diversity.

Bots survived: The number of bots that survive and reproduce each generation. Green racecar is the top bot of the last generation, Yellow racecars are survivors from previous generation, Red racecars are mutations of the Green and Yellow survivors.

Ticks per Generation: The game runs at 60 ticks per generation. This determines the number of ticks the simulation runs for initially.

Time Increase: This determines the number of ticks the simulation time increases by. The number of ticks per generation is equal to = Time Increase * Generations + Ticks per Generation.

Mutation Rate: Probability any given weight or bias is mutated. 0.5 -> 50% chance, 0.64 -> 64% chance, ect.

Mutation Range: The highest and lowest possible values the weight or bias could be mutated by. For example, if Mutation range is 0.75, a weight the value of 0.25 could become 1, -0.5, or anything in between

Rotation Punishment: A multiplier to determine the punishment for rotation. Fitness is calculated using average velocity, with points docked for high amounts of rotation.

Wall Punishment: Additional punishment for when a car touches wall. When set at zero, the only incentive for bots to avoid walls is is natural slowdown that comes with colliding.