Evolved neuralnet AI in CrystalSpace

The CEL demo application "evolvecraft" using a genetic algorithm (GA) to train a neural network (NN) to fly a hovercraft. This shows one generation very early on in the evolution. Notice how it mostly behaves quite stupidly. Tomorrow I'll post a video showing how it does after evolving several more generations.

NN: 4 inputs, 1 output, 2 hidden layers. Inputs are speed, and 3 "hitbeams" telling it the distance to any obstacle in front of it. Output simply turns it left or right. Activation function is tanh.

GA: Population size 20. Tournament selection. One-point crossover. Fitness function is distance travelled before it collides with the world.

http://www.crystalspace3d.org/

Category:

Entertainment

Tags:

• ai
• neural
• network
• genetic
• algorithm
• evolution
• crystal
• space
• game

[oktal3700]

I will post a better video some time soon. The fitness function just needs some work and then it should be quite good.

It's just a proof of concept anyway, for my work on the AI of the game Ecksdee (a.k.a XD), it's not intended to be totally awesome.

I am hampered by not having much time at the moment and very patchy internet access at home. Give me time...

[FireStdead]

Can't wait to see it :D I'm getting interested in these ANN stuff. I wish I could see some working code online especially the 'learning' part. I'm used to do some 'Flash developer' work and proper ANN can boost some Flash games. Is there open source ANN?

[oktal3700]

The software used in this video is open source C++ code, part of CEL (go to the CrystalSpace URL in this video's info text)

[FireStdead]

Actually it evolved a bit :D Stay in place and rotate - survival tactic!

[oktal3700]

Not really ;) The fitness function is a simple "distance traveled" so if it doesn't get very far then it's classed as unfit :P

[billamu]

Brrrmmm-crash... Brmm-crash... Brmm-crash... Brm-crash... Brmm-crash-Brmm-crash-Brmm-cra­sh.

Has it evolved?

[oktal3700]

Not much. Left to run overnight it can get to the edge of the level before crashing. That's as much as it can do since the fitness function is just the straight-line distance from the start position.

[03ricr]

@oktal3700

then its survival of the unfittest :)

[Curixq]

where is the new video? I would really want to see him. thx

[sistermiller]

cool

[fabsb0t]

now i have tried doing it myself i now realise how hard it is, well done!

sorry about my previous comment :P

[oktal3700]

not as good yet, but it will be. such is the nature of things