As a Formula 1 fan, I'm in awe of the drivers' ability to achieve such precision and tight margins on the track. I can't drive an F1 car (yet), but I can code one! So, I'm hand-drawing different maps inspired by the F1 calendar and letting my AI F1 racer take the wheel.
Implemented using NEAT (NeuroEvolution of Augmenting Topologies), NEAT is an algorithm for evolving artificial neural networks using:
- Genetic Encoding: Each neural network is encoded as a genome, which includes nodes and connections.
- Crossover and Mutation: Genomes can undergo crossover (combining parts of two genomes) and mutation (random changes) to produce offspring.
- Speciation: Genomes are grouped into species based on their structural similarity, which helps maintain diversity in the population.
- Innovation Numbers: Each new structure (node or connection) is assigned a unique innovation number, allowing for effective crossover between different topologies.
The neural network uses a radar system input which then passes through a hidden layer to 4 output neurons which correspond to Left, Right, Speed Up, Slow Down. For each action we implement a reward system based on the fitness metric. After each generation we keep the top 2 fittest cars from the last generation and evolve them until they reach max number of generations or complete 10 laps without crashing.
