NEW NEAT Hopper, 2004-01-09 (started 2002-10-17)

Hoppers can flip upside-down. This turns out to be a good tactic.

Here's a hopper that's lost it's stability.

Here's a hopper that's done fairly well.

Created by Mattias Fagerlund ( mattias@hypeskeptic.com ).

The fitness of a hopper is determined by how far it gets in a set amount of time. The programs that control the hopper are evolved neural networks, using a method called NEAT. I'm using a partial NEAT implementation, written in Delphi.

These hoppers are not identical to the GP Hoppers, though I hope to create a new GP Hopper program that can be compared to NEAT Hoppers. The differences are many and large;

• NEAT Hoppers have more inputs
• NEAT Hoppers use a different control mechanism
• NEAT Hoppers are expected to travel a shorter distance and
• the basic physical environment is different (still using ODE, though)

The shorter distance is in place because I want the hopper to succeed, so I can try different evolutionary strategies. The number of generations (on average) it takes to succeed is a good measurement for the quality of the method used. That quality is only really relevant to that particular problem, though. It's easy to over fit an evolutionary method.

The GP Hoppers basically try to travel as far as possible, the NEAT Hoppers try to travel 8,5 meters. A NEAT Hopper that succeeds in travelling 8,5 is considered a winner, and that run is terminated.

I'm assuming that NEAT is better at solving the problem than GP, but that remains to be tested under conditions where the results are actually comparable.

Download

Click here to download the the NEAT Hopper Size=811,30 kb, created 2004-08-19 21:53:37.

Click here to go back to the evolved page.