NEAT Demos
Watch how the evolution of increasingly
complex networks using NEAT results in more sophisticated behaviors in a robot
duel. View the neurons firing in real time as the robots compete.
Also see the
related publication.
See robots that were evolved to navigate a
simulated hallway using NEAT.
Neural networks are shown firing in real time as the robots move.
Includes entering a doorway and stopping before a doorway.
Watch NEAT neural networks
complexify as they approximate a difficult function. See how changing
structure contributes to function. Note: This experiment was coded
and designed by Mattias
Fagerlund, who has made some other NEAT demos available through his web
page.
Another excellent and entertaining demo of NEAT by Mattias Fagerlund. The hopper, which is like a robo-pogo-stick, learns to do things like travel forward, high jump, and maximize speed. It even can learn how to stand still and balance in an interesting demonstration reminiscent of pole balancing.
I suggest running it first with only the following options checked:
- Only draw better (only shows you when a champ improves)
- Real-time draw (draws at a real-world frame rate)
- Joint limits (doesnt let piece of the hopper move through each other)
- Foot trace (shows the path it took)
- Shadows (look nice)
To speed things up, I would also uncheck "multi-run," which doesn't seem necessary. (The program starts with this option checked by default)
The "Forward Distance" fitness measure is probably the best for
demonstration purposes, but "Standtime" is also interesting,
where the hopper learns to balance without falling. Also make
sure you can see both of the two windows. The second window
can show species breakdowns or the current champ's topology.
Real-time Interactive Neuroevolution Demo
"Peons," small video game characters that attempt to find goldmines and avoid an enemy, were evolved in real-time as the game progressed. See how they evolved to handle novel enemy strategies and goldmine placement scenarios. These demos are from older research that did not involve complexifying (NEAT) neural networks.