Racing game controlled by a neural network created by an evolutionary algorithm

First minute run of our evolved race driver. So far it seems like a pretty smart individual.
Take in mind, this driver started with no knowledge at al, we evolved a Neural Network (using NEAT) which controls this car. Through novelty search this is the best individual after 120 generations with 120 individuals per generation.

This is for our assignment in the Evolutionary algorithms course for the Master of the Artificial Intelligence programme at the University of Nijmegen.

I will post the best individual we find in the end.

Category:

Science & Technology

Tags:

• Evolutionary Algorithms
• TORCS
• race driver
• evolution
• artificial intelligence
• NEAT
• neural network
• racing
• car

[roleplayman]

what data do you feed to the network?

[jaapsant]

@roleplayman We had 25 input units, 19 of which were the distance sensors for every 10 degrees in the 0-180 degree half-circle in front of the car, 1 for the position of the car on the track, the angle relative to the track, speed in x-y-z direction, gear, revs, brakes.

This was one of the first finished evolved cars, after we had changed some things they got way better than this, perhaps i will post one of the final ones some time.

[cassiofoliveira]

I know its a litle bit late, but you should consider to reduce the amount of inputs. This way its going to evolve a lot faster... I used 5 distance sensors only, it worked just fine.

[mutehero7]

I would love to see this after maybe ten thousand generations :D

[roleplayman]

thanks very much for your reply , this will help me allot i am making a neural network in torcs as testing platform too, thanks.